Merge feat/calm-suggestions: drop no-op suggestions + suppress fickle re-edits

Claude-Session: https://claude.ai/code/session_016Yr6jELuRc7hyzYLccQKZd

internal/db/db.go

@@ -336,6 +336,35 @@ CREATE INDEX idx_vocab_due ON vocab_words(user_id, due_at);
 			name: "0007_vocab_definition",
 			stmt: `
 ALTER TABLE vocab_words ADD COLUMN definition TEXT NOT NULL DEFAULT '';
+`,
+		},
+		{
+			// Deterministic mechanics pass. Adds a 'mechanics' suggestion family for
+			// rule-based fixes (doubled words, spacing/punctuation, lowercase "i",
+			// curated confusables) detected in pure Go — no LLM. As with 0005, the
+			// `type` CHECK can't be ALTERed in place, so rebuild the table with the
+			// extended constraint, copy every row across, and recreate the index.
+			name: "0008_mechanics_suggestion_type",
+			stmt: `
+CREATE TABLE suggestions_new (
+    id           TEXT PRIMARY KEY DEFAULT (lower(hex(randomblob(16)))),
+    doc_id       TEXT NOT NULL REFERENCES documents(id) ON DELETE CASCADE,
+    from_pos     INTEGER NOT NULL,
+    to_pos       INTEGER NOT NULL,
+    original     TEXT NOT NULL,
+    replacement  TEXT NOT NULL,
+    explanation  TEXT NOT NULL,
+    type         TEXT NOT NULL CHECK(type IN ('grammar','phrasing','idiom','clarity','voice','collocation','mechanics')),
+    status       TEXT NOT NULL DEFAULT 'pending' CHECK(status IN ('pending','accepted','rejected')),
+    created_at   DATETIME DEFAULT CURRENT_TIMESTAMP
+);
+
+INSERT INTO suggestions_new (id, doc_id, from_pos, to_pos, original, replacement, explanation, type, status, created_at)
+SELECT id, doc_id, from_pos, to_pos, original, replacement, explanation, type, status, created_at FROM suggestions;
+
+DROP TABLE suggestions;
+ALTER TABLE suggestions_new RENAME TO suggestions;
+CREATE INDEX idx_suggestions_doc_id ON suggestions(doc_id);
 `,
 		},
 	}

internal/db/models.go

@@ -101,6 +101,7 @@ const (
 	SuggestionTypeClarity     = "clarity"
 	SuggestionTypeVoice       = "voice"
 	SuggestionTypeCollocation = "collocation"
+	SuggestionTypeMechanics   = "mechanics" // deterministic rule-based pass (no LLM)
 
 	SuggestionStatusPending  = "pending"
 	SuggestionStatusAccepted = "accepted"

internal/llm/checkpoint.go

@@ -89,6 +89,14 @@ func ParseCheckpoint(raw string) ([]RawSuggestion, error) {
 		if s.Original == "" {
 			continue
 		}
+		// Drop no-ops: the model sometimes "flags" a correct sentence and echoes
+		// it verbatim as the replacement, producing a card whose before/after are
+		// identical and whose explanation says it's already fine — a suggestion
+		// that suggests nothing. Awareness-only families (voice) carry an empty
+		// replacement, so this only ever fires on the edit families.
+		if s.Replacement != "" && s.Original == s.Replacement {
+			continue
+		}
 		out = append(out, s)
 	}
 	return out, nil

internal/llm/checkpoint_test.go

@@ -37,6 +37,25 @@ func TestParseCheckpoint(t *testing.T) {
 			raw:     "I could not find any issues!",
 			wantErr: true,
 		},
+		{
+			// A correct sentence echoed verbatim as its own replacement is a card
+			// that suggests nothing — dropped. The second item is a real edit.
+			name: "drops no-op where replacement equals original",
+			raw:  `{"suggestions":[{"original":"It wasn't the most graceful morning.","replacement":"It wasn't the most graceful morning.","explanation":"used correctly here","type":"idiom"},{"original":"teh","replacement":"the","explanation":"typo","type":"grammar"}]}`,
+			want: 1,
+		},
+		{
+			// Whitespace-only difference is still a no-op once both are trimmed.
+			name: "drops no-op after whitespace trim",
+			raw:  `{"suggestions":[{"original":"the cat sat","replacement":"  the cat sat  ","explanation":"already fine","type":"grammar"}]}`,
+			want: 0,
+		},
+		{
+			// Awareness-only voice flags carry an empty replacement and must survive.
+			name: "keeps awareness-only flag with empty replacement",
+			raw:  `{"suggestions":[{"original":"a long run-on passage","replacement":"","explanation":"this drifts from your voice","type":"voice"}]}`,
+			want: 1,
+		},
 		{
 			name: "braces inside string values",
 			raw:  `{"suggestions":[{"original":"use {x}","replacement":"use x","explanation":"drop the braces","type":"clarity"}]}`,

internal/suggestions/handlers.go

@@ -8,7 +8,9 @@ package suggestions
 import (
 	"context"
 	"database/sql"
+	"encoding/json"
 	"errors"
+	"io"
 	"net/http"
 	"strings"
 
@@ -46,6 +48,7 @@ func New(database *db.DB, client llm.LLMClient) *Handler {
 // the existing /api/docs router so they share its base path.
 func (h *Handler) RegisterDocRoutes(r chi.Router) {
 	r.Post("/{id}/check", h.check)
+	r.Post("/{id}/mechanics", h.mechanics)
 	r.Post("/{id}/voice", h.voice)
 	r.Post("/{id}/collocation", h.collocation)
 	r.Post("/{id}/rewrite", h.rewrite)
@@ -64,10 +67,118 @@ func (h *Handler) Routes() chi.Router {
 }
 
 // check runs a grammar checkpoint over the document. Fast, typing-cadence pass.
+// The deterministic mechanics family is owned by a separate pass (see mechanics),
+// detected client-side; a grammar checkpoint leaves those flags untouched.
 func (h *Handler) check(w http.ResponseWriter, r *http.Request) {
 	h.runPass(w, r, h.Limit, llm.RunCheckpoint, grammarScope)
 }
 
+// mechanicsFinding is one deterministic, rule-based fix detected client-side (see
+// web Companion/prose.ts). Detection lives in the frontend — the same rules that
+// power the companion's prose notes — so the server only persists these; it does
+// not compute them. Offsets are exact plaintext spans from the detector.
+type mechanicsFinding struct {
+	From        int    `json:"from"`
+	To          int    `json:"to"`
+	Original    string `json:"original"`
+	Replacement string `json:"replacement"`
+	Explanation string `json:"explanation"`
+}
+
+// maxMechanicsFindings caps a single submission so a runaway client can't flood
+// the table; far above any realistic count for one document.
+const maxMechanicsFindings = 500
+
+// mechanics persists the client-detected deterministic fixes as the 'mechanics'
+// family and returns the document's unified pending set. Free and not rate-
+// limited — it runs alongside the grammar checkpoint. It mirrors a single LLM
+// family: it replaces only the pending mechanics rows, honours actioned-
+// suppression, and (unlike the LLM passes) keeps the detector's exact offsets
+// rather than re-locating by string, which matters when the same word repeats.
+func (h *Handler) mechanics(w http.ResponseWriter, r *http.Request) {
+	docID := chi.URLParam(r, "id")
+
+	// Confirm the document exists (and is the local user's) for clean 404s.
+	var exists bool
+	err := h.DB.QueryRow(
+		`SELECT EXISTS(SELECT 1 FROM documents WHERE id = ? AND user_id = ?)`,
+		docID, db.LocalUserID,
+	).Scan(&exists)
+	if err != nil {
+		httputil.ServerError(w, err)
+		return
+	}
+	if !exists {
+		httputil.ErrorJSON(w, http.StatusNotFound, "document not found")
+		return
+	}
+
+	var body struct {
+		Findings []mechanicsFinding `json:"findings"`
+	}
+	if err := json.NewDecoder(io.LimitReader(r.Body, 1<<20)).Decode(&body); err != nil {
+		httputil.BadRequest(w, "invalid request body")
+		return
+	}
+	if len(body.Findings) > maxMechanicsFindings {
+		body.Findings = body.Findings[:maxMechanicsFindings]
+	}
+
+	if err := h.replaceMechanics(docID, body.Findings); err != nil {
+		httputil.ServerError(w, err)
+		return
+	}
+
+	out, err := h.fetchPending(docID)
+	if err != nil {
+		httputil.ServerError(w, err)
+		return
+	}
+	httputil.WriteJSON(w, http.StatusOK, out)
+}
+
+// replaceMechanics swaps the document's pending mechanics rows for the supplied
+// findings in one transaction, leaving the LLM families and actioned rows
+// untouched. Findings the user already accepted or dismissed are suppressed (the
+// detector has no memory between runs), and malformed spans are skipped.
+func (h *Handler) replaceMechanics(docID string, findings []mechanicsFinding) error {
+	tx, err := h.DB.Begin()
+	if err != nil {
+		return err
+	}
+	defer tx.Rollback()
+
+	if _, err := tx.Exec(
+		`DELETE FROM suggestions WHERE doc_id = ? AND status = ? AND type = ?`,
+		docID, db.SuggestionStatusPending, db.SuggestionTypeMechanics,
+	); err != nil {
+		return err
+	}
+
+	sup, err := buildSuppressor(tx, docID)
+	if err != nil {
+		return err
+	}
+
+	for _, f := range findings {
+		if f.From < 0 || f.To <= f.From || strings.TrimSpace(f.Original) == "" {
+			continue // malformed span — the client re-anchors by string anyway
+		}
+		if sup.suppressed(f.Original, f.Replacement) {
+			continue
+		}
+		if _, err := tx.Exec(
+			`INSERT INTO suggestions (doc_id, from_pos, to_pos, original, replacement, explanation, type)
+			 VALUES (?, ?, ?, ?, ?, ?, ?)`,
+			docID, f.From, f.To, f.Original, f.Replacement, f.Explanation, db.SuggestionTypeMechanics,
+		); err != nil {
+			return err
+		}
+	}
+
+	return tx.Commit()
+}
+
 // voice runs a Tier-1 voice-consistency pass over the whole document. Slow,
 // explicit-action pass; replaces only the pending voice flags.
 func (h *Handler) voice(w http.ResponseWriter, r *http.Request) {
@@ -163,9 +274,11 @@ type pendingScope struct {
 
 var (
 	// grammarScope owns the grammar/phrasing/idiom/clarity flags — everything but
-	// the explicit-action families (voice, collocation), which run on their own
-	// cadence and must survive a grammar checkpoint.
-	grammarScope = pendingScope{deleteWhere: "type NOT IN ('voice','collocation')", forceType: ""}
+	// the other self-owned families (voice, collocation, mechanics), which run on
+	// their own cadence/pass and must survive a grammar checkpoint. Notably the
+	// deterministic mechanics pass writes its rows in the same /check request just
+	// before this DELETE runs, so excluding it here is what keeps them alive.
+	grammarScope = pendingScope{deleteWhere: "type NOT IN ('voice','collocation','mechanics')", forceType: ""}
 	// voiceScope owns the voice flags only.
 	voiceScope = pendingScope{deleteWhere: "type = 'voice'", forceType: db.SuggestionTypeVoice}
 	// collocationScope owns the collocation flags only.
@@ -176,13 +289,12 @@ var (
 // fresh batch in a single transaction. Accepted/rejected suggestions and the
 // other family's pending rows are left untouched.
 //
-// Suggestions the user already accepted or dismissed are suppressed from the
-// fresh batch: the model has no memory between passes, so without this it would
-// re-propose the identical edit on the very next checkpoint — re-nagging a
-// sentence the user already resolved. This matters most when an accept silently
-// no-ops (the `original` text couldn't be anchored in the editor, so the doc
-// never changed): the sentence is unaltered, yet the user shouldn't see the same
-// card again.
+// Suggestions touching a sentence the user already settled are suppressed from
+// the fresh batch (see suppressor): not just the identical edit re-proposed, but
+// reversals and re-polishing of the model's own just-accepted output — the
+// "fickle, keeps going back and forth on a few sentences" behavior. The model has
+// no memory between passes, so without this it re-opens resolved sentences every
+// checkpoint.
 func (h *Handler) replacePending(docID, contentText string, raw []llm.RawSuggestion, scope pendingScope) error {
 	tx, err := h.DB.Begin()
 	if err != nil {
@@ -197,13 +309,13 @@ func (h *Handler) replacePending(docID, contentText string, raw []llm.RawSuggest
 		return err
 	}
 
-	actioned, err := actionedKeys(tx, docID)
+	sup, err := buildSuppressor(tx, docID)
 	if err != nil {
 		return err
 	}
 
 	for _, s := range raw {
-		if _, seen := actioned[suggestionKey(s.Original, s.Replacement)]; seen {
+		if sup.suppressed(s.Original, s.Replacement) {
 			continue
 		}
 		typ := scope.forceType
@@ -223,37 +335,131 @@ func (h *Handler) replacePending(docID, contentText string, raw []llm.RawSuggest
 	return tx.Commit()
 }
 
-// actionedKeys returns the set of original→replacement keys the user has already
-// accepted or rejected for this document, so a fresh checkpoint can skip
-// re-proposing them. Keyed on the trimmed original+replacement pair, matching the
-// normalization ParseCheckpoint applies, so a genuinely different edit on the same
-// sentence is not suppressed.
-func actionedKeys(tx *sql.Tx, docID string) (map[string]struct{}, error) {
+// dedupQuoteReplacer folds every straight/curly single- and double-quote variant
+// (and backtick/acute accent) onto one canonical character. The editor and the
+// model both rewrite quotes between passes — a sentence accepted with "…" comes
+// back flagged with '…' — so without folding, byte-identical text reads as a
+// different edit and the suppression below misses it. (This normalization is for
+// dedup ONLY; the frontend still anchors on the verbatim `original`.)
+var dedupQuoteReplacer = strings.NewReplacer(
+	"‘", "'", "’", "'", "‚", "'", "‛", "'", // single curly
+	"“", "'", "”", "'", "„", "'", "″", "'", // double curly
+	"\"", "'", "`", "'", "´", "'", // straight double, backtick, acute
+)
+
+// normalizeForDedup canonicalizes a string for suppression comparisons: quotes
+// folded (above) and runs of whitespace collapsed to single spaces (so a reflowed
+// paragraph still matches). Used only to decide what to suppress, never to alter
+// stored or rendered text.
+func normalizeForDedup(s string) string {
+	return strings.Join(strings.Fields(dedupQuoteReplacer.Replace(s)), " ")
+}
+
+// suppressor decides which fresh suggestions to drop because the user has already
+// settled the sentence they touch. The model has no memory between passes, so on
+// every checkpoint it re-examines the current text and proposes edits — including
+// ones that re-open a sentence the user already resolved. Three families of those
+// are suppressed (all compared under normalizeForDedup):
+//
+//   - pairs: the identical edit, re-proposed verbatim (the original "accept it,
+//     then it nags again" case; also covers a silent no-op accept that left the
+//     text unchanged).
+//   - actionedOrig: any edit whose original is a span the user already accepted or
+//     dismissed an edit on — "you already decided about this exact sentence."
+//   - acceptedRepl: any edit whose original is text the user accepted AS a
+//     replacement — i.e. the model re-touching its own just-accepted output, which
+//     is how the reversals and endless re-polishing arise (accept "due to the
+//     rain", next pass proposes changing "due to the rain" back). Guarded to
+//     multi-word spans so word-level fixes aren't swept up as collateral.
+//   - acceptedReplList holds the same accepted replacements for a containment
+//     check: the model evades the exact acceptedRepl match by re-flagging a
+//     *sub-clause* of an accepted sentence (flag "she was…due to the rain" instead
+//     of the whole sentence). When one of the new original / an accepted
+//     replacement contains the other and the shorter side is substantial
+//     (>= minContainWords words), it's the same settled span and is dropped.
+type suppressor struct {
+	pairs            map[string]struct{}
+	actionedOrig     map[string]struct{}
+	acceptedRepl     map[string]struct{}
+	acceptedReplList []string
+}
+
+// minContainWords is the floor for the containment check: the shorter of the two
+// spans must be at least this many words before a substring relationship counts
+// as "the same settled text." High enough that an incidental common phrase ("the
+// rain") can't suppress an unrelated sentence, low enough to catch a re-flagged
+// clause.
+const minContainWords = 4
+
+// suppressed reports whether a fresh suggestion should be dropped as already
+// settled. An empty original is never suppressed here (it can't anchor anyway and
+// is dropped upstream).
+func (s suppressor) suppressed(original, replacement string) bool {
+	o := normalizeForDedup(original)
+	if o == "" {
+		return false
+	}
+	if _, ok := s.pairs[o+"\x00"+normalizeForDedup(replacement)]; ok {
+		return true
+	}
+	if _, ok := s.actionedOrig[o]; ok {
+		return true
+	}
+	if strings.ContainsRune(o, ' ') {
+		if _, ok := s.acceptedRepl[o]; ok {
+			return true
+		}
+	}
+	// Containment: the model re-flagged a sub-clause of (or a window around) an
+	// accepted span. Suppress when one contains the other and the shorter side is
+	// a substantial multi-word run.
+	for _, r := range s.acceptedReplList {
+		shorter, longer := o, r
+		if len(r) < len(o) {
+			shorter, longer = r, o
+		}
+		if len(strings.Fields(shorter)) >= minContainWords && strings.Contains(longer, shorter) {
+			return true
+		}
+	}
+	return false
+}
+
+// buildSuppressor loads the document's accepted/rejected edits and indexes them
+// into the three suppression families described on suppressor.
+func buildSuppressor(tx *sql.Tx, docID string) (suppressor, error) {
 	rows, err := tx.Query(
-		`SELECT original, replacement FROM suggestions
+		`SELECT original, replacement, status FROM suggestions
 		  WHERE doc_id = ? AND status IN (?, ?)`,
 		docID, db.SuggestionStatusAccepted, db.SuggestionStatusRejected,
 	)
 	if err != nil {
-		return nil, err
+		return suppressor{}, err
 	}
 	defer rows.Close()
 
-	keys := make(map[string]struct{})
-	for rows.Next() {
-		var original, replacement string
-		if err := rows.Scan(&original, &replacement); err != nil {
-			return nil, err
-		}
-		keys[suggestionKey(original, replacement)] = struct{}{}
+	s := suppressor{
+		pairs:        make(map[string]struct{}),
+		actionedOrig: make(map[string]struct{}),
+		acceptedRepl: make(map[string]struct{}),
 	}
-	return keys, rows.Err()
-}
-
-// suggestionKey is the dedup identity for a suggestion: its trimmed original and
-// replacement text. Two suggestions with the same key are "the same edit."
-func suggestionKey(original, replacement string) string {
-	return strings.TrimSpace(original) + "\x00" + strings.TrimSpace(replacement)
+	for rows.Next() {
+		var original, replacement, status string
+		if err := rows.Scan(&original, &replacement, &status); err != nil {
+			return suppressor{}, err
+		}
+		o := normalizeForDedup(original)
+		r := normalizeForDedup(replacement)
+		s.pairs[o+"\x00"+r] = struct{}{}
+		if o != "" {
+			s.actionedOrig[o] = struct{}{}
+		}
+		if status == db.SuggestionStatusAccepted && r != "" {
+			s.acceptedRepl[r] = struct{}{}
+			s.acceptedReplList = append(s.acceptedReplList, r)
+		}
+	}
+	return s, rows.Err()
 }
 
 // listForDoc returns the document's current pending suggestions (used when the
@@ -291,7 +497,51 @@ func (h *Handler) fetchPending(docID string) ([]db.Suggestion, error) {
 		}
 		out = append(out, s)
 	}
-	return out, rows.Err()
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	return dedupeSpans(out), nil
+}
+
+// dedupeSpans resolves collisions between the deterministic mechanics family and
+// the LLM families: when a mechanics finding and an LLM suggestion fight over the
+// same characters, mechanics wins and the LLM card is dropped. Its span is exact
+// (the detector matched it), whereas the LLM positions are only advisory
+// (re-anchored by string at render), so the precise fix should own the span.
+//
+// This deliberately does NOT dedupe LLM-vs-LLM overlaps: voice (awareness-only,
+// no replacement) and collocation legitimately co-occupy the same span, and that
+// is intended. Suggestions that never anchored (from_pos < 0) occupy no real span
+// and are always kept.
+func dedupeSpans(in []db.Suggestion) []db.Suggestion {
+	type span struct{ from, to int }
+	var claimed []span
+	for _, s := range in {
+		if s.Type == db.SuggestionTypeMechanics && s.FromPos >= 0 {
+			claimed = append(claimed, span{s.FromPos, s.ToPos})
+		}
+	}
+	if len(claimed) == 0 {
+		return in
+	}
+
+	out := make([]db.Suggestion, 0, len(in))
+	for _, s := range in {
+		if s.Type != db.SuggestionTypeMechanics && s.FromPos >= 0 {
+			overlaps := false
+			for _, sp := range claimed {
+				if s.FromPos < sp.to && sp.from < s.ToPos {
+					overlaps = true
+					break
+				}
+			}
+			if overlaps {
+				continue // an exact mechanics fix owns these characters
+			}
+		}
+		out = append(out, s)
+	}
+	return out
 }
 
 // accept marks a suggestion accepted (the client applies the replacement text).

internal/suggestions/handlers_test.go

@@ -165,6 +165,48 @@ func TestResolvedSuggestionsNotReproposed(t *testing.T) {
 	}
 }
 
+// TestFickleEditsSuppressed proves the suppressor kills the "keeps going back and
+// forth on a few sentences" behavior seen live on the Missing Key doc: a later
+// pass that (a) reverses an edit the user just accepted — even with the editor's
+// double→single quote churn that defeats a byte-exact match — or (b) re-polishes
+// the model's own accepted output. A genuinely new edit on a fresh sentence still
+// survives.
+func TestFickleEditsSuppressed(t *testing.T) {
+	client := &stubClient{response: `{"suggestions":[
+		{"original":"He left \"early,\" because of the rain.","replacement":"He left \"early,\" due to the rain.","explanation":"smoother","type":"phrasing"},
+		{"original":"The cat always have a calm face.","replacement":"The cat always has a calm face.","explanation":"agreement","type":"grammar"}
+	]}`}
+	srv, docID, h := newTestServer(t, client)
+	h.Limit = llm.NewRateLimiter(0)
+
+	rec := do(t, srv, http.MethodPost, "/docs/"+docID+"/check", "")
+	var got []db.Suggestion
+	_ = json.Unmarshal(rec.Body.Bytes(), &got)
+	if len(got) != 2 {
+		t.Fatalf("first pass: want 2, got %d", len(got))
+	}
+	for _, s := range got {
+		do(t, srv, http.MethodPost, "/suggestions/"+s.ID+"/accept", "")
+	}
+
+	// Reversal of the first accept (note the " → ' quote churn) and a re-polish of
+	// the second accept must both be dropped; only the unrelated edit survives.
+	client.response = `{"suggestions":[
+		{"original":"He left 'early,' due to the rain.","replacement":"He left 'early,' because of the rain.","explanation":"reverts the accepted edit","type":"phrasing"},
+		{"original":"The cat always has a calm face.","replacement":"The cat always seems to have a calm face.","explanation":"re-polishes accepted output","type":"phrasing"},
+		{"original":"due to the rain","replacement":"because of the rain","explanation":"sub-clause reversal of the accepted span","type":"phrasing"},
+		{"original":"She drived home.","replacement":"She drove home.","explanation":"past tense","type":"grammar"}
+	]}`
+	rec = do(t, srv, http.MethodPost, "/docs/"+docID+"/check", "")
+	var again []db.Suggestion
+	if err := json.Unmarshal(rec.Body.Bytes(), &again); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	if len(again) != 1 || again[0].Original != "She drived home." {
+		t.Fatalf("want only the new edit to survive, got %d: %+v", len(again), again)
+	}
+}
+
 // TestVoicePassCoexists proves the voice pass and the grammar checkpoint own
 // disjoint pending families: running one never wipes the other's flags, and
 // each endpoint returns the unified pending set.
@@ -316,3 +358,116 @@ func TestCheckpointRateLimit(t *testing.T) {
 		t.Fatalf("rate limit should suppress the second model call, got %d calls", client.calls)
 	}
 }
+
+// postMechanics submits a deterministic-findings batch (as the client's prose
+// detector would) and returns the unified pending set.
+func postMechanics(t *testing.T, srv http.Handler, docID, findingsJSON string) []db.Suggestion {
+	t.Helper()
+	rec := do(t, srv, http.MethodPost, "/docs/"+docID+"/mechanics", `{"findings":`+findingsJSON+`}`)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("mechanics: code=%d body=%s", rec.Code, rec.Body)
+	}
+	var got []db.Suggestion
+	if err := json.Unmarshal(rec.Body.Bytes(), &got); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	return got
+}
+
+// TestMechanicsPersistsFindings proves the endpoint persists client-detected
+// findings as a 'mechanics' family with the exact offsets the client supplied,
+// and that a later grammar checkpoint leaves them untouched (own family).
+func TestMechanicsPersistsFindings(t *testing.T) {
+	client := &stubClient{response: `{"suggestions":[]}`}
+	srv, docID, _ := newTestServer(t, client)
+
+	got := postMechanics(t, srv, docID, `[
+		{"from":0,"to":1,"original":"i","replacement":"I","explanation":"capitalize I"},
+		{"from":6,"to":13,"original":"the the","replacement":"the","explanation":"doubled word"}
+	]`)
+	var mech []db.Suggestion
+	for _, s := range got {
+		if s.Type == db.SuggestionTypeMechanics {
+			mech = append(mech, s)
+		}
+	}
+	if len(mech) != 2 {
+		t.Fatalf("want 2 mechanics findings, got %d: %+v", len(mech), got)
+	}
+	// The client's exact offsets are preserved verbatim.
+	for _, s := range mech {
+		if s.Original == "the the" && (s.FromPos != 6 || s.ToPos != 13) {
+			t.Errorf("offsets not preserved: %+v", s)
+		}
+	}
+
+	// A grammar checkpoint must not wipe the mechanics family.
+	rec := do(t, srv, http.MethodPost, "/docs/"+docID+"/check", "")
+	if rec.Code != http.StatusOK {
+		t.Fatalf("check: code=%d body=%s", rec.Code, rec.Body)
+	}
+	var after []db.Suggestion
+	_ = json.Unmarshal(rec.Body.Bytes(), &after)
+	count := 0
+	for _, s := range after {
+		if s.Type == db.SuggestionTypeMechanics {
+			count++
+		}
+	}
+	if count != 2 {
+		t.Fatalf("grammar checkpoint disturbed the mechanics family: got %d, want 2", count)
+	}
+}
+
+// TestMechanicsWinsSpanCollision proves that when a mechanics finding and an LLM
+// grammar suggestion claim overlapping characters, the LLM card is dropped from
+// the response and the exact mechanics fix owns the span.
+func TestMechanicsWinsSpanCollision(t *testing.T) {
+	// LLM grammar suggestion covers "the the cat", overlapping the mechanics
+	// "the the" finding at [0,7].
+	client := &stubClient{response: `{"suggestions":[
+		{"original":"the the cat","replacement":"the cat","explanation":"wordy","type":"grammar"}
+	]}`}
+	srv, docID, h := newTestServer(t, client)
+	if _, err := h.DB.Exec(`UPDATE documents SET content_text = ? WHERE id = ?`, "the the cat sat", docID); err != nil {
+		t.Fatalf("set content: %v", err)
+	}
+
+	// Persist the mechanics finding, then run the grammar pass.
+	postMechanics(t, srv, docID, `[{"from":0,"to":7,"original":"the the","replacement":"the","explanation":"doubled word"}]`)
+	rec := do(t, srv, http.MethodPost, "/docs/"+docID+"/check", "")
+	if rec.Code != http.StatusOK {
+		t.Fatalf("check: code=%d body=%s", rec.Code, rec.Body)
+	}
+	var got []db.Suggestion
+	if err := json.Unmarshal(rec.Body.Bytes(), &got); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	for _, s := range got {
+		if s.Type == db.SuggestionTypeGrammar {
+			t.Fatalf("overlapping grammar card should have been dropped in favor of mechanics, got %+v", got)
+		}
+	}
+	if len(got) != 1 || got[0].Type != db.SuggestionTypeMechanics {
+		t.Fatalf("want sole mechanics finding, got %+v", got)
+	}
+}
+
+// TestMechanicsActionedSuppression proves a dismissed mechanics fix is not
+// re-proposed on the next submission of the same finding.
+func TestMechanicsActionedSuppression(t *testing.T) {
+	client := &stubClient{response: `{"suggestions":[]}`}
+	srv, docID, _ := newTestServer(t, client)
+
+	findings := `[{"from":6,"to":13,"original":"the the","replacement":"the","explanation":"doubled word"}]`
+	got := postMechanics(t, srv, docID, findings)
+	if len(got) != 1 {
+		t.Fatalf("want 1 finding, got %+v", got)
+	}
+	// Dismiss it, then resubmit the same finding — it must stay suppressed.
+	do(t, srv, http.MethodPost, "/suggestions/"+got[0].ID+"/dismiss", "")
+	again := postMechanics(t, srv, docID, findings)
+	if len(again) != 0 {
+		t.Fatalf("dismissed mechanics fix should not reappear, got %+v", again)
+	}
+}

web/src/App.tsx

@@ -274,7 +274,7 @@ export default function App() {
       if (currentDoc) {
         patchSummary(currentDoc.id, { word_count: change.word_count })
         schedule(change)
-        scheduleCheckpoint()
+        scheduleCheckpoint(change.content_text)
       }
     },
     [currentDoc, patchSummary, schedule, scheduleCheckpoint],

web/src/api/client.ts

@@ -76,7 +76,7 @@ export interface Gloss {
   gloss: string
 }
 
-export type SuggestionType = 'grammar' | 'phrasing' | 'idiom' | 'clarity' | 'voice' | 'collocation'
+export type SuggestionType = 'grammar' | 'phrasing' | 'idiom' | 'clarity' | 'voice' | 'collocation' | 'mechanics'
 
 // One word in the vocabulary garden: a looked-up word with its gloss/phonetic,
 // the sentence it was met in, and its spaced-repetition state. `reps` drives how
@@ -135,6 +135,17 @@ export interface Suggestion {
   created_at: string
 }
 
+// A deterministic, rule-based fix detected client-side (see Companion/prose.ts).
+// The frontend owns mechanics detection; the backend only persists these as the
+// 'mechanics' suggestion family. Spans are exact plaintext offsets.
+export interface MechanicsFinding {
+  from: number
+  to: number
+  original: string
+  replacement: string
+  explanation: string
+}
+
 async function req<T>(path: string, init?: RequestInit): Promise<T> {
   const res = await fetch(`/api${path}`, {
     headers: { 'Content-Type': 'application/json' },
@@ -168,6 +179,14 @@ export const api = {
   // word pairings ("do a decision" → "make a decision"). Returns the unified
   // pending set too. Rate-limited per document server-side.
   collocationDoc: (id: string) => req<Suggestion[]>(`/docs/${id}/collocation`, { method: 'POST' }),
+  // Mechanics pass: persist the client-detected deterministic fixes as the
+  // 'mechanics' family and return the unified pending set. Not rate-limited (it's
+  // free, local detection); runs alongside the grammar checkpoint.
+  submitMechanics: (id: string, findings: MechanicsFinding[]) =>
+    req<Suggestion[]>(`/docs/${id}/mechanics`, {
+      method: 'POST',
+      body: JSON.stringify({ findings }),
+    }),
   // Pending suggestions for a doc, loaded when the editor opens it.
   listSuggestions: (id: string) => req<Suggestion[]>(`/docs/${id}/suggestions`),
   acceptSuggestion: (id: string) =>

web/src/components/Companion/prose.test.ts

@@ -248,3 +248,119 @@ describe('capitalization basics', () => {
     expectRule('The sun set slowly. it was a beautiful evening down by the river.', 'cap-sentence')
   })
 })
+
+import { mechanicsFindings } from './prose'
+
+// mechanicsFindings is the applyable subset feeding the suggestion cards. Each
+// finding must carry an EXACT span (text.slice(from,to) === original) and a
+// replacement that actually changes the text — these become one-click edits.
+describe('mechanicsFindings — applyable fixes', () => {
+  // Every finding's span must be exact, regardless of which rule produced it.
+  function expectExactSpans(text: string) {
+    const found = mechanicsFindings(text)
+    for (const f of found) {
+      expect(text.slice(f.from, f.to), `span for ${JSON.stringify(f)}`).toBe(f.original)
+      expect(f.replacement).not.toBe(f.original)
+    }
+    return found
+  }
+
+  const one = (text: string, rulePred: (f: ReturnType<typeof mechanicsFindings>[number]) => boolean) =>
+    expectExactSpans(text).find(rulePred)
+
+  it('doubled word → single word, exact span', () => {
+    const f = one('I saw the the cat in the garden today.', (f) => f.original === 'the the')
+    expect(f?.replacement).toBe('the')
+  })
+
+  it('finds BOTH doublings with distinct spans', () => {
+    const found = mechanicsFindings('the the dog and the the cat ran around the yard.')
+      .filter((f) => f.original === 'the the')
+    expect(found).toHaveLength(2)
+    expect(found[0].from).not.toBe(found[1].from)
+  })
+
+  // Spans are widened to a distinctive phrase so they re-anchor by string in the
+  // editor (a bare "a"/"i"/" is" would match the wrong spot). The replacement
+  // carries the whole corrected phrase.
+
+  it('lowercase standalone i is awareness-only (no card)', () => {
+    // It still flags as a companion bubble (see analyzeProse), but a single
+    // character can't be anchored, so it must NOT become a card.
+    expect(mechanicsFindings('Yesterday i walked to the store and came home.').length).toBe(0)
+    expect(rulesFor('Yesterday i walked to the store and came home.')).toContain('cap-i')
+  })
+
+  it('a → an: spans the whole "a <noun>" phrase', () => {
+    const f = one('She found a umbrella under the old wooden table.', (f) => f.original === 'a umbrella')
+    expect(f?.replacement).toBe('an umbrella')
+  })
+
+  it('uncountable plural → singular', () => {
+    const f = one('She gave me many informations about the new job.', (f) => f.original === 'informations')
+    expect(f?.replacement).toBe('information')
+  })
+
+  it('lowercase proper noun → capitalized', () => {
+    const f = one('I am learning english during my free time this year.', (f) => f.original === 'english')
+    expect(f?.replacement).toBe('English')
+  })
+
+  it('subject-verb agreement → corrects the verb, spans subject+verb', () => {
+    const f = one('She have three cats and a dog at her house.', (f) => f.original === 'She have')
+    expect(f?.replacement).toBe('She has')
+  })
+
+  it('singular noun after a number → plural, spans number+noun', () => {
+    const f = one('I bought five apple at the market this morning.', (f) => f.original === 'five apple')
+    expect(f?.replacement).toBe('five apples')
+  })
+
+  it('comparative + then → than, spans the phrase', () => {
+    const f = one('This book is better then the one I read last week.', (f) => f.original === 'better then')
+    expect(f?.replacement).toBe('better than')
+  })
+
+  it('space before punctuation is removed (spans the word)', () => {
+    const f = one('I love it , it makes me very happy indeed.', (f) => f.original === 'it ,')
+    expect(f?.replacement).toBe('it,')
+  })
+
+  it('missing space after a comma is added (spans both words)', () => {
+    const f = one('I bought apples,bananas and pears at the store.', (f) => f.original === 'apples,bananas')
+    expect(f?.replacement).toBe('apples, bananas')
+  })
+
+  it('stacked determiner drops the article', () => {
+    const f = one('She left the my book on the kitchen table again.', (f) => f.original === 'the my')
+    expect(f?.replacement).toBe('my')
+  })
+
+  it('there is + plural → there are, spans the phrase', () => {
+    const f = one('There is many people waiting outside in the cold.', (f) => f.original === 'There is many')
+    expect(f?.replacement).toBe('There are many')
+  })
+
+  it("it's own → its own, spans the phrase", () => {
+    const f = one("The cat licked it's own paw very slowly today.", (f) => f.original === "it's own")
+    expect(f?.replacement).toBe('its own')
+  })
+
+  it("its a → it's a, spans the phrase", () => {
+    const f = one('I think its a wonderful day to go outside now.', (f) => f.original === 'its a')
+    expect(f?.replacement).toBe("it's a")
+  })
+
+  // Awareness-only rules never produce cards — they stay companion bubbles.
+  it('run-ons and splices produce NO card findings', () => {
+    const runOn = 'I went to the market and I bought some apples and then I saw my friend and we talked for a while and after that we went to the cafe and ordered coffee and cake together.'
+    expect(mechanicsFindings(runOn).some((f) => f.original.length > 30)).toBe(false)
+    const splice = 'I finished the report, I sent it to my boss right away today.'
+    // no card spans the comma-join; only (if any) small word-level fixes
+    expect(mechanicsFindings(splice).every((f) => f.to - f.from < 12)).toBe(true)
+  })
+
+  it('clean prose yields no findings', () => {
+    expect(mechanicsFindings('The garden was quiet and the rain fell softly on the leaves.')).toEqual([])
+  })
+})

web/src/components/Companion/prose.ts

@@ -8,6 +8,14 @@
 // Each finding becomes a Mandarin-first bubble Line (see tips.ts), often quoting
 // a short slice of her own sentence so the advice clearly belongs to *this*
 // paragraph and not a generic tip jar.
+//
+// Two consumers share these rules (one source of truth, no duplication):
+//   • the companion bubbles — awareness notes, shown one at a time (analyzeProse)
+//   • the suggestion cards  — applyable one-click fixes (mechanicsFindings)
+// A rule is "applyable" when it can name an exact span and a single replacement;
+// it attaches a `fix` and feeds the cards. Awareness-only rules (run-ons, comma
+// splices, …) carry no `fix` and stay companion bubbles. The companion hides
+// fix-bearing hints so the same span never appears as both a bubble and a card.
 
 import type { Line } from './tips'
 
@@ -17,6 +25,28 @@ export interface ProseHint extends Line {
   id: string
   // Rule family, used to vary advice (don't show the same kind twice in a row).
   rule: string
+  // Present only on *applyable* findings: the exact span and the text to swap in.
+  // These become suggestion cards; the companion skips them (see useCompanion).
+  fix?: Fix
+}
+
+// An exact, one-click edit: replace text[from:to] with `replacement`.
+export interface Fix {
+  from: number
+  to: number
+  replacement: string
+}
+
+// A deterministic suggestion-card finding, derived from an applyable hint. Mirrors
+// the backend's card shape (original/replacement/explanation + span) so the card
+// pipeline can persist it as the 'mechanics' family. `explanation` is the English
+// note; the card's own translate action renders Mandarin on demand.
+export interface MechanicsFinding {
+  from: number
+  to: number
+  original: string
+  replacement: string
+  explanation: string
 }
 
 // ── small text helpers ──────────────────────────────────────────────────────
@@ -48,11 +78,23 @@ function key(s: string): string {
   return s.toLowerCase().replace(/\s+/g, ' ').trim().slice(0, 48)
 }
 
+// Capitalize `repl`'s first letter when `original` started uppercase, so fixing a
+// sentence-initial word doesn't quietly lowercase the line.
+function matchCase(original: string, repl: string): string {
+  if (!original || !repl) return repl
+  const c = original[0]
+  if (c >= 'A' && c <= 'Z') return repl[0].toUpperCase() + repl.slice(1)
+  return repl
+}
+
 // ── individual rules ────────────────────────────────────────────────────────
-// Each rule pushes at most a couple of findings; the caller shows one at a time.
+// Each rule pushes its findings; awareness rules cap themselves to a couple so a
+// single sentence can't flood the bubble, while applyable rules report every
+// occurrence (each becomes its own card). Applyable rules attach a `fix`.
 
 // Run-on / overly long sentences — the single most common readability problem
 // for ESL writers, who often chain clauses that a period would serve better.
+// Awareness-only: there is no single mechanical fix for "split this sentence".
 function runOns(text: string, out: ProseHint[]) {
   let found = 0
   for (const s of sentences(text)) {
@@ -110,6 +152,7 @@ const INTRO_LEAD = new Set([
 // approximate that by requiring the lead to be ≥3 words and not begin with an
 // introductory word — short or transition-led leads are intro phrases, not
 // spliced clauses. Precision over recall: a wrong nudge costs more than a miss.
+// Awareness-only: the fix is ambiguous (period vs. a joining word).
 function commaSplices(text: string, out: ProseHint[]) {
   let found = 0
   for (const s of sentences(text)) {
@@ -137,6 +180,7 @@ function commaSplices(text: string, out: ProseHint[]) {
 // Unclear antecedent — a sentence that opens with This/That/These/Those riding
 // straight into a verb, with no noun naming what it points back to. Only flag
 // when there's a prior sentence (so an antecedent is actually in question).
+// Awareness-only: naming the referent needs the writer.
 const ANTECEDENT_RE =
   /^(This|That|These|Those)\s+(is|are|was|were|will|would|can|could|should|makes?|made|means?|shows?|showed|gives?|gave|causes?|caused|creates?|created|leads?|led|results?|happens?|happened|helps?|helped)\b/
 
@@ -160,7 +204,8 @@ function antecedents(text: string, out: ProseHint[]) {
 
 // Oxford comma — a 3+ item list ending “… X and Y” with no comma before the
 // conjunction. Guarded by a clause-starter stoplist so we don't mistake a
-// compound clause (“…, but she and I…”) for a list.
+// compound clause (“…, but she and I…”) for a list. Awareness-only: inserting
+// the serial comma is borderline-stylistic, so we nudge rather than auto-edit.
 const OXFORD_RE = /(\w+),\s+([\w'’-]+(?:\s+[\w'’-]+){0,2})\s+(and|or)\s+[\w'’-]+/g
 const CLAUSE_STARTERS = new Set([
   'but', 'so', 'because', 'which', 'who', 'that', 'when', 'while', 'if',
@@ -187,81 +232,32 @@ function oxford(text: string, out: ProseHint[]) {
   }
 }
 
-// a vs. an — chosen by the *sound* of the next word. We work only on lowercase
-// words (sidestepping proper nouns / acronyms) and keep sound-exception lists.
-const A_BEFORE_VOWEL_RE = /\ba\s+([aeiou][a-z]+)\b/g
-// Vowel-spelled but consonant-sounding → “a” is correct, don't flag.
-const CONSONANT_SOUND_RE = /^(uni|use|usu|util|euro?|eul|ewe|once|one|ubiqu|unanim)/
-const AN_BEFORE_CONSONANT_RE = /\ban\s+([b-df-hj-np-tv-z][a-z]+)\b/g
-// Consonant-spelled but vowel-sounding (silent h) → “an” is correct, don't flag.
-const VOWEL_SOUND_RE = /^(hour|honest|honou?r|heir|homage)/
+// A transition word opening a sentence with no comma after it (“However we…”).
+// Limited to conjunctive adverbs that strongly want the comma — sequence words
+// like “Then/First/Finally” are left out (their comma is optional). Awareness-
+// only: it fires per-sentence (offsets are sentence-relative), so we nudge.
+const INTRO_RE =
+  /^(However|Therefore|Moreover|Furthermore|Nevertheless|Nonetheless|Meanwhile|Consequently|In addition|In conclusion|As a result|On the other hand|For example|For instance)\s+[A-Za-z]/
 
-function articles(text: string, out: ProseHint[]) {
-  let m: RegExpExecArray | null
-  let found = 0
-  A_BEFORE_VOWEL_RE.lastIndex = 0
-  while ((m = A_BEFORE_VOWEL_RE.exec(text)) && found < 1) {
-    if (CONSONANT_SOUND_RE.test(m[1])) continue
-    found++
-    out.push({
-      id: `article-an:${key(m[1])}`,
-      rule: 'article',
-      zh: `元音开头的词前用 “an”：“an ${m[1]}”。`,
-      en: `Before a vowel sound, use “an”: “an ${m[1]}”.`,
-    })
-  }
-  AN_BEFORE_CONSONANT_RE.lastIndex = 0
-  while ((m = AN_BEFORE_CONSONANT_RE.exec(text)) && found < 2) {
-    if (VOWEL_SOUND_RE.test(m[1])) continue
-    found++
-    out.push({
-      id: `article-a:${key(m[1])}`,
-      rule: 'article',
-      zh: `辅音开头的词前用 “a”：“a ${m[1]}”。`,
-      en: `Before a consonant sound, use “a”: “a ${m[1]}”.`,
-    })
-  }
-}
-
-// Doubled word — “the the”, “is is”. Excludes the handful of doublings that can
-// be legitimate (“the fact that that happened”, “she had had enough”).
-const DOUBLE_RE = /\b([A-Za-z]+)\s+\1\b/gi
-const LEGIT_DOUBLES = new Set(['that', 'had'])
-
-function doubles(text: string, out: ProseHint[]) {
-  let m: RegExpExecArray | null
-  let found = 0
-  DOUBLE_RE.lastIndex = 0
-  while ((m = DOUBLE_RE.exec(text)) && found < 1) {
-    const w = m[1].toLowerCase()
-    if (LEGIT_DOUBLES.has(w)) continue
-    found++
-    out.push({
-      id: `double:${key(m[0])}`,
-      rule: 'double',
-      zh: `“${m[1]}” 好像写了两遍，检查一下哦。`,
-      en: `“${m[1]} ${m[1]}” — looks like a word got doubled.`,
-    })
-  }
-}
-
-// Lowercase standalone “I”. Conservative: only when bounded by spaces / line
-// edge, to avoid tangling with “i.e.” and the like.
-const LOWER_I_RE = /(?:^|\s)i(?=\s|$)/m
-
-function pronounI(text: string, out: ProseHint[]) {
-  if (LOWER_I_RE.test(text)) {
-    out.push({
-      id: 'cap-i',
-      rule: 'cap-i',
-      zh: '英文里的 “I”（我）任何时候都要大写哦。',
-      en: 'In English, “I” is always written as a capital letter.',
-    })
+function introComma(text: string, out: ProseHint[]) {
+  for (const s of sentences(text)) {
+    const m = s.match(INTRO_RE)
+    if (m) {
+      out.push({
+        id: `introcomma:${m[1].toLowerCase()}`,
+        rule: 'introcomma',
+        zh: `开头的过渡词后面加个逗号：“${m[1]}, …”。`,
+        en: `Put a comma after the opening transition: “${m[1]}, …”.`,
+      })
+      break
+    }
   }
 }
 
 // Sentence not starting with a capital. We look after a terminator, and ignore
 // CJK sentences (she writes Mandarin too — those don't take Latin capitals).
+// Awareness-only: a bare “. x” also matches abbreviations (“U.S. then”), so we
+// nudge rather than auto-capitalize.
 const LOWER_START_RE = /[.!?]\s+([a-z])/
 
 function sentenceCaps(text: string, out: ProseHint[]) {
@@ -275,40 +271,148 @@ function sentenceCaps(text: string, out: ProseHint[]) {
   }
 }
 
-// A stray space *before* punctuation (a common habit carried from CJK spacing).
-const SPACE_BEFORE_PUNCT_RE = /[A-Za-z] +([,;:!?]|\.(?!\.))/
+// ── applyable rules (also feed the suggestion cards) ─────────────────────────
 
-function spaceBeforePunct(text: string, out: ProseHint[]) {
-  if (SPACE_BEFORE_PUNCT_RE.test(text)) {
+// Doubled word — “the the”, “is is”. Excludes the handful of doublings that can
+// be legitimate (“the fact that that happened”, “she had had enough”).
+const DOUBLE_RE = /\b([A-Za-z]+)(\s+)\1\b/gi
+const LEGIT_DOUBLES = new Set(['that', 'had'])
+
+function doubles(text: string, out: ProseHint[]) {
+  let m: RegExpExecArray | null
+  DOUBLE_RE.lastIndex = 0
+  while ((m = DOUBLE_RE.exec(text))) {
+    const w = m[1].toLowerCase()
+    if (LEGIT_DOUBLES.has(w)) continue
+    const from = m.index
+    const to = from + m[0].length
     out.push({
-      id: 'space-punct',
-      rule: 'space-punct',
-      zh: '标点前面不用空格，逗号、句号紧跟在前一个词后面就好。',
-      en: 'No space before punctuation — it tucks right against the word.',
+      id: `double:${from}:${key(m[0])}`,
+      rule: 'double',
+      zh: `“${m[1]}” 好像写了两遍，检查一下哦。`,
+      en: `“${m[1]} ${m[1]}” — looks like a word got doubled.`,
+      fix: { from, to, replacement: m[1] },
     })
   }
 }
 
-// ── Chinese-L1 interference rules (Tier 1) ──────────────────────────────────
-// Mandarin lacks plural inflection, articles, and verb agreement, so these
-// patterns are the predictable places that grammar "leaks" into her English.
-// All are closed-list or strong-signal to keep false positives near zero.
+// Lowercase standalone “I”. Conservative: only when bounded by spaces / line
+// edge, to avoid tangling with “i.e.” and the like. Awareness-only: a single
+// character can't be re-anchored unambiguously by string (every word holds an
+// “i”), so this stays a companion nudge rather than a one-click card.
+const LOWER_I_RE = /(^|[^\p{L}\p{N}_])i(?=$|[^\p{L}\p{N}_])/mu
+
+function pronounI(text: string, out: ProseHint[]) {
+  const m = LOWER_I_RE.exec(text)
+  LOWER_I_RE.lastIndex = 0
+  if (!m) return
+  const at = m.index + m[1].length
+  if (text[at + 1] === '.') return // “i.e.” etc. — precision over recall
+  out.push({
+    id: 'cap-i',
+    rule: 'cap-i',
+    zh: '英文里的 “I”（我）任何时候都要大写哦。',
+    en: 'In English, “I” is always written as a capital letter.',
+  })
+}
+
+// A stray space *before* punctuation (a common habit carried from CJK spacing).
+const SPACE_BEFORE_PUNCT_RE = /([A-Za-z]+) +([,;:!?]|\.(?!\.))/g
+
+function spaceBeforePunct(text: string, out: ProseHint[]) {
+  let m: RegExpExecArray | null
+  SPACE_BEFORE_PUNCT_RE.lastIndex = 0
+  while ((m = SPACE_BEFORE_PUNCT_RE.exec(text))) {
+    const from = m.index
+    const to = from + m[0].length
+    out.push({
+      id: `space-punct:${from}`,
+      rule: 'space-punct',
+      zh: '标点前面不用空格，逗号、句号紧跟在前一个词后面就好。',
+      en: 'No space before punctuation — it tucks right against the word.',
+      fix: { from, to, replacement: m[1] + m[2] },
+    })
+  }
+}
+
+// Missing space *after* a comma or sentence period (“apple,banana”, “end.Next”).
+// The comma case requires letters on both sides so numbers like 1,000 are safe;
+// the period case requires a real word boundary so “e.g.”/“U.S.” are skipped.
+const NO_SPACE_COMMA_RE = /([A-Za-z]+,)([A-Za-z]+)/g
+const NO_SPACE_PERIOD_RE = /([a-z]{2,}\.)([A-Z][a-z]+)/g
+
+function spaceAfterPunct(text: string, out: ProseHint[]) {
+  for (const re of [NO_SPACE_COMMA_RE, NO_SPACE_PERIOD_RE]) {
+    let m: RegExpExecArray | null
+    re.lastIndex = 0
+    while ((m = re.exec(text))) {
+      const from = m.index
+      const to = from + m[0].length
+      out.push({
+        id: `space-after:${from}`,
+        rule: 'space-after',
+        zh: '逗号、句号后面要空一格，再接下一个词。',
+        en: 'Add a space after a comma or period before the next word.',
+        fix: { from, to, replacement: `${m[1]} ${m[2]}` },
+      })
+    }
+  }
+}
+
+// a vs. an — chosen by the *sound* of the next word. We work only on lowercase
+// words (sidestepping proper nouns / acronyms) and keep sound-exception lists.
+const A_BEFORE_VOWEL_RE = /\b(a)\s+([aeiou][a-z]+)\b/g
+// Vowel-spelled but consonant-sounding → “a” is correct, don't flag.
+const CONSONANT_SOUND_RE = /^(uni|use|usu|util|euro?|eul|ewe|once|one|ubiqu|unanim)/
+const AN_BEFORE_CONSONANT_RE = /\b(an)\s+([b-df-hj-np-tv-z][a-z]+)\b/g
+// Consonant-spelled but vowel-sounding (silent h) → “an” is correct, don't flag.
+const VOWEL_SOUND_RE = /^(hour|honest|honou?r|heir|homage)/
+
+function articles(text: string, out: ProseHint[]) {
+  let m: RegExpExecArray | null
+  A_BEFORE_VOWEL_RE.lastIndex = 0
+  while ((m = A_BEFORE_VOWEL_RE.exec(text))) {
+    if (CONSONANT_SOUND_RE.test(m[2])) continue
+    // Span the whole "a <noun>" so the original is a distinctive, anchorable
+    // phrase (a bare "a" matches everywhere). Replacement swaps only the article.
+    out.push({
+      id: `article-an:${m.index}`,
+      rule: 'article',
+      zh: `元音开头的词前用 “an”：“an ${m[2]}”。`,
+      en: `Before a vowel sound, use “an”: “an ${m[2]}”.`,
+      fix: { from: m.index, to: m.index + m[0].length, replacement: matchCase(m[1], 'an') + m[0].slice(m[1].length) },
+    })
+  }
+  AN_BEFORE_CONSONANT_RE.lastIndex = 0
+  while ((m = AN_BEFORE_CONSONANT_RE.exec(text))) {
+    if (VOWEL_SOUND_RE.test(m[2])) continue
+    out.push({
+      id: `article-a:${m.index}`,
+      rule: 'article',
+      zh: `辅音开头的词前用 “a”：“a ${m[2]}”。`,
+      en: `Before a consonant sound, use “a”: “a ${m[2]}”.`,
+      fix: { from: m.index, to: m.index + m[0].length, replacement: matchCase(m[1], 'a') + m[0].slice(m[1].length) },
+    })
+  }
+}
 
 // Mass nouns that Mandarin speakers very commonly pluralize. These words are
 // essentially never valid with a trailing “s”, so the list is its own guard.
 const UNCOUNTABLE_PLURAL_RE =
-  /\b(informations|advices|knowledges|equipments|furnitures|homeworks|softwares|hardwares|luggages|baggages|sceneries|machineries)\b/i
+  /\b(informations|advices|knowledges|equipments|furnitures|homeworks|softwares|hardwares|luggages|baggages|sceneries|machineries)\b/gi
 
 function uncountables(text: string, out: ProseHint[]) {
-  const m = UNCOUNTABLE_PLURAL_RE.exec(text)
+  let m: RegExpExecArray | null
   UNCOUNTABLE_PLURAL_RE.lastIndex = 0
-  if (m) {
-    const singular = m[1].replace(/s$/i, '')
+  while ((m = UNCOUNTABLE_PLURAL_RE.exec(text))) {
+    const word = m[1]
+    const singular = word.replace(/s$/i, '')
     out.push({
-      id: `uncountable:${m[1].toLowerCase()}`,
+      id: `uncountable:${m.index}`,
       rule: 'uncountable',
-      zh: `“${m[1]}” 是不可数名词，不用加 s，写 “${singular}” 就好。`,
-      en: `“${m[1]}” is uncountable — drop the “s”: just “${singular}”.`,
+      zh: `“${word}” 是不可数名词，不用加 s，写 “${singular}” 就好。`,
+      en: `“${word}” is uncountable — drop the “s”: just “${singular}”.`,
+      fix: { from: m.index, to: m.index + word.length, replacement: singular },
     })
   }
 }
@@ -324,17 +428,20 @@ const PROPER_NOUNS =
   'italian|russian|portuguese|arabic|vietnamese|thai|american|british|canadian|' +
   'australian|mexican|brazilian|european'
 // Case-sensitive (lowercase-only) so already-capitalized words aren't flagged.
-const PROPER_NOUN_RE = new RegExp(`\\b(${PROPER_NOUNS})\\b`)
+const PROPER_NOUN_RE = new RegExp(`\\b(${PROPER_NOUNS})\\b`, 'g')
 
 function properCaps(text: string, out: ProseHint[]) {
-  const m = PROPER_NOUN_RE.exec(text)
-  if (m) {
-    const fixed = m[1][0].toUpperCase() + m[1].slice(1)
+  let m: RegExpExecArray | null
+  PROPER_NOUN_RE.lastIndex = 0
+  while ((m = PROPER_NOUN_RE.exec(text))) {
+    const word = m[1]
+    const fixed = word[0].toUpperCase() + word.slice(1)
     out.push({
-      id: `propercap:${m[1]}`,
+      id: `propercap:${m.index}`,
       rule: 'propercap',
       zh: `语言、国籍、星期和月份在英文里要大写：“${fixed}”。`,
       en: `Languages, days, and months are capitalized in English: “${fixed}”.`,
+      fix: { from: m.index, to: m.index + word.length, replacement: fixed },
     })
   }
 }
@@ -367,17 +474,20 @@ const CAUSATIVE = new Set([
 
 function subjectVerbAgreement(text: string, out: ProseHint[]) {
   let m: RegExpExecArray | null
-  let found = 0
   SVA_RE.lastIndex = 0
-  while ((m = SVA_RE.exec(text)) && found < 2) {
+  while ((m = SVA_RE.exec(text))) {
     if (m[1] && CAUSATIVE.has(m[1].toLowerCase())) continue
-    found++
-    const fixed = third(m[3])
+    const verb = m[3]
+    const fixed = third(verb)
+    // Span the whole match (subject + verb, plus any captured lead) so the
+    // original anchors; the replacement corrects only the trailing verb.
+    const head = m[0].slice(0, m[0].length - verb.length)
     out.push({
-      id: `sva:${m[2].toLowerCase()}:${m[3].toLowerCase()}`,
+      id: `sva:${m.index}`,
       rule: 'sva',
       zh: `主语是 he/she/it 时，动词要加 -s：“${m[2]} ${fixed}”。`,
       en: `After he/she/it the verb takes “-s”: “${m[2]} ${fixed}”.`,
+      fix: { from: m.index, to: m.index + m[0].length, replacement: head + matchCase(verb, fixed) },
     })
   }
 }
@@ -397,35 +507,40 @@ const NON_S_PLURAL = new Set([
 
 function pluralAfterNumber(text: string, out: ProseHint[]) {
   let m: RegExpExecArray | null
-  let found = 0
   NUMBER_NOUN_RE.lastIndex = 0
-  while ((m = NUMBER_NOUN_RE.exec(text)) && found < 1) {
+  while ((m = NUMBER_NOUN_RE.exec(text))) {
     const noun = m[2].toLowerCase()
     if (noun.endsWith('s') || NON_S_PLURAL.has(noun)) continue
-    found++
+    // Span "<number> <noun>" so the original anchors; append “s” to the noun.
     out.push({
-      id: `plural:${key(m[0])}`,
+      id: `plural:${m.index}`,
       rule: 'plural',
       zh: `“${m[1]}” 后面的名词要用复数：“${m[1]} ${m[2]}s”。`,
       en: `After “${m[1]}”, the noun is plural: “${m[1]} ${m[2]}s”.`,
+      fix: { from: m.index, to: m.index + m[0].length, replacement: `${m[0]}s` },
     })
   }
 }
 
 // Two stacked determiners (“the my book”, “a the”) — Mandarin uses a bare
-// possessive, so the article often gets stacked on top. One of the two must go.
+// possessive, so the article often gets stacked on top. The article goes.
 const DOUBLE_DET_RE =
-  /\b(a|an|the)\s+(a|an|the|my|your|his|her|its|our|their)\b/gi
+  /\b(a|an|the)(\s+)(a|an|the|my|your|his|her|its|our|their)\b/gi
 
 function doubleDeterminer(text: string, out: ProseHint[]) {
-  const m = DOUBLE_DET_RE.exec(text)
+  let m: RegExpExecArray | null
   DOUBLE_DET_RE.lastIndex = 0
-  if (m) {
+  while ((m = DOUBLE_DET_RE.exec(text))) {
+    // Two *identical* words ("the the") are a doubled word, not stacked
+    // determiners — leave that to the `doubles` rule so we don't double-flag.
+    if (m[1].toLowerCase() === m[3].toLowerCase()) continue
     out.push({
-      id: `doubledet:${key(m[0])}`,
+      id: `doubledet:${m.index}`,
       rule: 'doubledet',
-      zh: `“${m[1]} ${m[2]}” 用了两个限定词，留一个就好（比如去掉 “${m[1]}”）。`,
-      en: `“${m[1]} ${m[2]}” stacks two determiners — keep just one.`,
+      zh: `“${m[1]} ${m[3]}” 用了两个限定词，留一个就好（比如去掉 “${m[1]}”）。`,
+      en: `“${m[1]} ${m[3]}” stacks two determiners — keep just one.`,
+      // Drop the article (m[1]); keep the second determiner, casing preserved.
+      fix: { from: m.index, to: m.index + m[0].length, replacement: matchCase(m[1], m[3]) },
     })
   }
 }
@@ -434,46 +549,52 @@ function doubleDeterminer(text: string, out: ProseHint[]) {
 // on clearly-plural cues (skip “some/any”, which are fine with singular mass
 // nouns: “there is some water”).
 const THERE_IS_RE =
-  /\bthere(?:\s+is|'s|’s)\s+(many|several|numerous|various|few|two|three|four|five|six|seven|eight|nine|ten)\b/i
+  /\bthere(\s+is|'s|’s)\s+(many|several|numerous|various|few|two|three|four|five|six|seven|eight|nine|ten)\b/gi
 
 function thereIsPlural(text: string, out: ProseHint[]) {
-  const m = THERE_IS_RE.exec(text)
+  let m: RegExpExecArray | null
   THERE_IS_RE.lastIndex = 0
-  if (m) {
+  while ((m = THERE_IS_RE.exec(text))) {
+    // Span the whole "there is <plural>" phrase so it anchors; swap the verb part
+    // (m[1]: “ is” / “'s”) for “ are”, preserving the “there” casing and the cue.
+    const replacement = m[0].slice(0, 5) + ' are' + m[0].slice(5 + m[1].length)
     out.push({
-      id: `thereis:${m[1].toLowerCase()}`,
+      id: `thereis:${m.index}`,
       rule: 'thereis',
-      zh: `后面是复数时用 “there are”：“there are ${m[1]}…”。`,
-      en: `With a plural, use “there are”: “there are ${m[1]}…”.`,
+      zh: `后面是复数时用 “there are”：“there are ${m[2]}…”。`,
+      en: `With a plural, use “there are”: “there are ${m[2]}…”.`,
+      fix: { from: m.index, to: m.index + m[0].length, replacement },
     })
   }
 }
 
-// ── confusables (Tier 2) ─────────────────────────────────────────────────────
-
 // its / it's — only the two unambiguous directions: “it's own” (always its own)
 // and “its a/an” (the possessive can't take an article → it's a/an).
-const ITS_OWN_RE = /\bit's\s+own\b/i
-const ITS_ARTICLE_RE = /\bits\s+(a|an)\b/i
+const ITS_OWN_RE = /\b(it's|it’s)\s+own\b/gi
+const ITS_ARTICLE_RE = /\bits\s+(a|an)\b/gi
 
 function itsConfusion(text: string, out: ProseHint[]) {
-  if (ITS_OWN_RE.test(text)) {
+  let m: RegExpExecArray | null
+  ITS_OWN_RE.lastIndex = 0
+  while ((m = ITS_OWN_RE.exec(text))) {
+    // Span "it's own" so it anchors; swap the leading token to the possessive.
     out.push({
-      id: 'its-own',
+      id: `its-own:${m.index}`,
       rule: 'its',
       zh: '“it’s” = “it is”；表示“它的”要用 “its”，所以是 “its own”。',
       en: '“it’s” means “it is” — the possessive is “its”: “its own”.',
+      fix: { from: m.index, to: m.index + m[0].length, replacement: matchCase(m[1], 'its') + m[0].slice(m[1].length) },
     })
-    return
   }
-  const m = ITS_ARTICLE_RE.exec(text)
   ITS_ARTICLE_RE.lastIndex = 0
-  if (m) {
+  while ((m = ITS_ARTICLE_RE.exec(text))) {
+    // Span "its a/an" so it anchors; swap "its" → "it's".
     out.push({
-      id: 'its-article',
+      id: `its-article:${m.index}`,
       rule: 'its',
       zh: `这里应该是 “it’s ${m[1]}”（it is），“its” 是“它的”。`,
       en: `Here it should be “it’s ${m[1]}” (it is); “its” means belonging to it.`,
+      fix: { from: m.index, to: m.index + m[0].length, replacement: matchCase(m[0][0], "it's") + m[0].slice(3) },
     })
   }
 }
@@ -485,57 +606,20 @@ const COMPARATIVES =
   'better|more|less|rather|worse|greater|other|older|younger|bigger|smaller|' +
   'larger|faster|slower|higher|lower|cheaper|stronger|weaker|easier|harder|' +
   'earlier|later|sooner|longer|shorter|taller|richer|poorer|happier|safer'
-const THAN_THEN_RE = new RegExp(`\\b(${COMPARATIVES})\\s+then\\b`, 'i')
+const THAN_THEN_RE = new RegExp(`\\b(${COMPARATIVES})(\\s+)then\\b`, 'gi')
 
 function thanThen(text: string, out: ProseHint[]) {
-  const m = THAN_THEN_RE.exec(text)
-  if (m) {
+  let m: RegExpExecArray | null
+  THAN_THEN_RE.lastIndex = 0
+  while ((m = THAN_THEN_RE.exec(text))) {
+    // Span "<comparative> then" so it anchors; swap the trailing “then” → “than”.
+    const thenFrom = m.index + m[0].length - 4
     out.push({
-      id: `than:${m[1].toLowerCase()}`,
+      id: `than:${m.index}`,
       rule: 'than',
       zh: `比较的时候用 “than”，不是 “then”：“${m[1]} than”。`,
       en: `For comparisons use “than”, not “then”: “${m[1]} than”.`,
-    })
-  }
-}
-
-// A transition word opening a sentence with no comma after it (“However we…”).
-// Limited to conjunctive adverbs that strongly want the comma — sequence words
-// like “Then/First/Finally” are left out (their comma is optional).
-const INTRO_RE =
-  /^(However|Therefore|Moreover|Furthermore|Nevertheless|Nonetheless|Meanwhile|Consequently|In addition|In conclusion|As a result|On the other hand|For example|For instance)\s+[A-Za-z]/
-
-function introComma(text: string, out: ProseHint[]) {
-  let found = false
-  for (const s of sentences(text)) {
-    const m = s.match(INTRO_RE)
-    if (m) {
-      out.push({
-        id: `introcomma:${m[1].toLowerCase()}`,
-        rule: 'introcomma',
-        zh: `开头的过渡词后面加个逗号：“${m[1]}, …”。`,
-        en: `Put a comma after the opening transition: “${m[1]}, …”.`,
-      })
-      found = true
-      break
-    }
-  }
-  return found
-}
-
-// Missing space *after* a comma or sentence period (“apple,banana”, “end.Next”).
-// The comma case requires letters on both sides so numbers like 1,000 are safe;
-// the period case requires a real word boundary so “e.g.”/“U.S.” are skipped.
-const NO_SPACE_COMMA_RE = /[A-Za-z],[A-Za-z]/
-const NO_SPACE_PERIOD_RE = /[a-z]{2,}\.[A-Z][a-z]/
-
-function spaceAfterPunct(text: string, out: ProseHint[]) {
-  if (NO_SPACE_COMMA_RE.test(text) || NO_SPACE_PERIOD_RE.test(text)) {
-    out.push({
-      id: 'space-after',
-      rule: 'space-after',
-      zh: '逗号、句号后面要空一格，再接下一个词。',
-      en: 'Add a space after a comma or period before the next word.',
+      fix: { from: m.index, to: m.index + m[0].length, replacement: m[0].slice(0, m[0].length - 4) + matchCase(text[thenFrom], 'than') },
     })
   }
 }
@@ -568,6 +652,8 @@ const RULES: Array<(text: string, out: ProseHint[]) => void> = [
 
 // analyzeProse returns context-aware hints, highest-priority first. It bails on
 // text too short to advise on (mid-thought drafts shouldn't get picked apart).
+// Hints that carry a `fix` are applyable (they also surface as suggestion cards);
+// the companion filters those out so a span isn't both a bubble and a card.
 export function analyzeProse(text: string): ProseHint[] {
   const englishWords = text.match(ENGLISH_WORD_RE)?.length ?? 0
   if (englishWords < 8) return []
@@ -575,3 +661,35 @@ export function analyzeProse(text: string): ProseHint[] {
   for (const rule of RULES) rule(text, out)
   return out
 }
+
+// mechanicsFindings returns every applyable deterministic fix in the text, as
+// suggestion-card findings with exact spans. No word-count floor: a doubled word
+// or a stray lowercase “i” is worth fixing even in a short draft, the way a
+// spell-checker would. The card pipeline persists these as the 'mechanics'
+// family; collisions with the LLM cards are resolved server-side (mechanics
+// wins, since its span is exact).
+export function mechanicsFindings(text: string): MechanicsFinding[] {
+  const hints: ProseHint[] = []
+  for (const rule of RULES) rule(text, hints)
+  const found: MechanicsFinding[] = []
+  for (const h of hints) {
+    if (!h.fix) continue
+    const { from, to, replacement } = h.fix
+    const original = text.slice(from, to)
+    if (!original || original === replacement) continue
+    found.push({ from, to, original, replacement, explanation: h.en })
+  }
+  // Two rules can occasionally claim overlapping spans (e.g. a doubled word that
+  // also reads as stacked determiners). Resolve to one card per stretch of text:
+  // earlier start wins, ties broken by the longer span. The backend resolves the
+  // separate mechanics-vs-LLM collisions; this handles mechanics-vs-mechanics.
+  found.sort((a, b) => (a.from !== b.from ? a.from - b.from : b.to - a.to))
+  const out: MechanicsFinding[] = []
+  let lastEnd = -1
+  for (const f of found) {
+    if (f.from < lastEnd) continue
+    out.push(f)
+    lastEnd = f.to
+  }
+  return out
+}

web/src/components/Companion/useCompanion.ts

@@ -127,7 +127,10 @@ export function useCompanion({ wordCount, saveStatus, llmDown, editTick, acceptT
   // every finding was shown recently. Avoids repeating a finding or the same
   // rule family back-to-back so the companion never feels like a broken record.
   const nextTip = useCallback((): Bubble => {
-    const hints = analyzeProse(textRef.current)
+    // Applyable hints (those with a `fix`) surface as one-click suggestion cards,
+    // so the companion skips them — a span shouldn't be both a bubble and a card.
+    // What's left is the awareness notes (run-ons, splices, …) the kitten alone gives.
+    const hints = analyzeProse(textRef.current).filter((h) => !h.fix)
     const fresh = hints.find(
       (h) => !recentHints.current.includes(h.id) && h.rule !== lastRule.current,
     )

web/src/components/Editor/AskPetal.tsx

@@ -34,9 +34,12 @@ export function AskPetal({ suggestionId, explanation }: Props) {
     if (el) el.scrollTop = el.scrollHeight
   }, [messages])
 
-  // Focus the input when the panel opens.
+  // Focus the input when the panel opens. preventScroll: the card is already on
+  // screen as an absolutely-positioned overlay, and a default focus() would make
+  // the browser scroll its ancestor to "reveal" the input — jumping the document
+  // to the top.
   useEffect(() => {
-    inputRef.current?.focus()
+    inputRef.current?.focus({ preventScroll: true })
   }, [])
 
   // Fetch the Chinese translation of the explanation to seed the first bubble.
@@ -95,7 +98,7 @@ export function AskPetal({ suggestionId, explanation }: Props) {
       console.error('Ask Petal chat failed:', err)
     } finally {
       setStreaming(false)
-      inputRef.current?.focus()
+      inputRef.current?.focus({ preventScroll: true })
     }
   }
 

web/src/components/Editor/EditorCore.tsx

@@ -580,9 +580,37 @@ export function EditorCore({
     [onDismiss, closeCard],
   )
 
-  // Click a red-underlined word to open its spelling popover, anchored under the
-  // word. posAtCoords→wordAt resolves the exact PM span (robust to the same
-  // misspelling appearing elsewhere); nspell supplies the corrections.
+  // openMisspellAt resolves the word at a document position and, if the checker
+  // flags it as misspelled, opens the spelling popover anchored under the word.
+  // Returns true if it opened a card. We resolve the word from coordinates and
+  // the checker rather than from the `.petal-misspelling` DOM span: clicking a
+  // word moves the caret into it, which fires the decoration rebuild that
+  // deliberately un-underlines the caret word (see SpellCheck's caret exemption),
+  // so by the time a click/contextmenu handler runs the span is already gone.
+  const openMisspellAt = useCallback(
+    (pos: number): boolean => {
+      if (!editor || !spellChecker) return false
+      const range = wordAt(editor.state.doc, pos)
+      if (!range || spellChecker.correct(range.word)) return false
+      const wrapper = wrapperRef.current
+      if (!wrapper) return false
+      const start = editor.view.coordsAtPos(range.from)
+      const end = editor.view.coordsAtPos(range.to)
+      const wrapRect = wrapper.getBoundingClientRect()
+      const cardWidth = 240
+      const left = Math.max(0, Math.min(start.left - wrapRect.left, wrapper.clientWidth - cardWidth))
+      const top = end.bottom - wrapRect.top + 6
+      // Opening a spelling popover supersedes any AI-suggestion or lookup card.
+      closeCard()
+      setWordInfo(null)
+      setMisspell({ ...range, suggestions: spellChecker.suggest(range.word), top, left })
+      return true
+    },
+    [editor, spellChecker, closeCard],
+  )
+
+  // Click a misspelled word to open its spelling popover, anchored under the
+  // word; nspell supplies the corrections.
   //
   // Tapping an AI-suggestion highlight also opens its card here — on touch there's
   // no hover, so the tap is the only way in (mouse users still get hover).
@@ -603,26 +631,16 @@ export function EditorCore({
         }
         return
       }
-      if (!editor || !spellChecker) return
-      const target = (e.target as HTMLElement).closest('.petal-misspelling') as HTMLElement | null
-      if (!target) return
-      const wrapper = wrapperRef.current
-      if (!wrapper) return
+      if (!editor) return
+      // Only act on clicks in the editor text itself — the floating cards/popovers
+      // are children of this same wrapper, and a click on one shouldn't resolve a
+      // (hidden) word behind it.
+      if (!(e.target as HTMLElement).closest('.petal-prose')) return
       const coords = editor.view.posAtCoords({ left: e.clientX, top: e.clientY })
       if (!coords) return
-      const range = wordAt(editor.state.doc, coords.pos)
-      if (!range) return
-      const elRect = target.getBoundingClientRect()
-      const wrapRect = wrapper.getBoundingClientRect()
-      const cardWidth = 240
-      const left = Math.max(0, Math.min(elRect.left - wrapRect.left, wrapper.clientWidth - cardWidth))
-      const top = elRect.bottom - wrapRect.top + 6
-      // Opening a spelling popover supersedes any AI-suggestion hover card.
-      closeCard()
-      setWordInfo(null)
-      setMisspell({ ...range, suggestions: spellChecker.suggest(range.word), top, left })
+      openMisspellAt(coords.pos)
     },
-    [editor, spellChecker, closeCard, openCardFor, railEnabled],
+    [editor, openMisspellAt, openCardFor, railEnabled],
   )
 
   const replaceMisspelling = useCallback(
@@ -757,8 +775,10 @@ export function EditorCore({
       .catch((err) => console.error('vocab save failed', err))
   }, [wordInfo, exampleAt, docId])
 
-  // Right-click a word to look it up. Right-clicking off any word falls through
-  // to the native menu (so copy/paste-by-menu still works — see the Selection fix).
+  // Right-click a misspelled word for spelling corrections (the familiar "did you
+  // mean" gesture); right-click any other word to look it up. Right-clicking off
+  // any word falls through to the native menu (so copy/paste-by-menu still works
+  // — see the Selection fix).
   const handleContextMenu = useCallback(
     (e: React.MouseEvent) => {
       if (!editor) return
@@ -766,9 +786,11 @@ export function EditorCore({
       if (!coords) return
       if (!wordAt(editor.state.doc, coords.pos)) return
       e.preventDefault()
+      // A misspelled word offers corrections first; otherwise look it up.
+      if (openMisspellAt(coords.pos)) return
       openWordLookup(coords.pos)
     },
-    [editor, openWordLookup],
+    [editor, openMisspellAt, openWordLookup],
   )
 
   // Touch has no hover or right-click, so a long-press (~500ms without moving)

web/src/components/Editor/suggestionMeta.ts

@@ -10,4 +10,5 @@ export const TYPE_META: Record<SuggestionType, { color: string; label: string }>
   clarity: { color: 'var(--color-sky)', label: 'Clarity' },
   voice: { color: 'var(--color-honey)', label: 'Voice' },
   collocation: { color: 'var(--color-blossom)', label: 'Word pairing' },
+  mechanics: { color: 'var(--color-sage)', label: 'Tidy-up' },
 }

web/src/hooks/useCheckpoint.ts

@@ -1,5 +1,6 @@
 import { useCallback, useEffect, useRef, useState } from 'react'
 import { api, type Suggestion } from '../api/client'
+import { mechanicsFindings } from '../components/Companion/prose'
 
 const DEBOUNCE_MS = 4000
 
@@ -27,6 +28,13 @@ export function useCheckpoint(docId: string | null) {
   const docIdRef = useRef(docId)
   docIdRef.current = docId
 
+  // Latest plaintext, kept fresh by schedule(), so the deterministic mechanics
+  // pass (detected client-side, see prose.ts) reads the current document when the
+  // debounce fires — no need to re-thread text through every call. null until the
+  // first edit of the current doc, so a tone-only check before any edit doesn't
+  // submit an empty batch and wipe the doc's persisted mechanics rows.
+  const latestTextRef = useRef<string | null>(null)
+
   // Token to discard responses from a doc we've since navigated away from.
   const runRef = useRef(0)
 
@@ -46,6 +54,21 @@ export function useCheckpoint(docId: string | null) {
     setChecking(true)
     let retrying = false
     try {
+      // Deterministic mechanics first: detect client-side and persist as the
+      // 'mechanics' family before the grammar pass, so the checkpoint's unified
+      // response already carries them. Best-effort and only on the initial try —
+      // a grammar retry shouldn't re-submit unchanged findings. A mechanics
+      // failure must not block the grammar pass.
+      if (attempt === 0 && latestTextRef.current !== null) {
+        try {
+          // Render the mechanics fixes immediately — they're instant (no LLM), so
+          // the unified set they return shouldn't wait on the slow grammar pass.
+          const withMech = await api.submitMechanics(id, mechanicsFindings(latestTextRef.current))
+          if (run === runRef.current && id === docIdRef.current) setSuggestions(withMech)
+        } catch (err) {
+          console.error('mechanics submit failed', err)
+        }
+      }
       const fresh = await api.checkDoc(id)
       if (run === runRef.current && id === docIdRef.current) {
         setSuggestions(fresh)
@@ -116,8 +139,11 @@ export function useCheckpoint(docId: string | null) {
     [runExplicitPass],
   )
 
-  // Call on every edit; schedules a check 4s after typing settles.
-  const schedule = useCallback(() => {
+  // Call on every edit; schedules a check 4s after typing settles. Pass the
+  // current plaintext so the mechanics pass sees the latest document; omit it
+  // (e.g. a tone-only change) to reuse the last text.
+  const schedule = useCallback((text?: string) => {
+    if (text !== undefined) latestTextRef.current = text
     clearTimeout(debounceRef.current)
     clearTimeout(retryRef.current) // a fresh edit supersedes any queued retry
     debounceRef.current = setTimeout(() => void runCheck(), DEBOUNCE_MS)
@@ -129,6 +155,7 @@ export function useCheckpoint(docId: string | null) {
     clearTimeout(debounceRef.current)
     clearTimeout(retryRef.current)
     runRef.current++
+    latestTextRef.current = null // unknown until the new doc's first edit
     setSuggestions([])
     setChecking(false)
     setVoicing(false)

web/src/index.css

@@ -21,6 +21,7 @@
   --color-sky: #A8CCE8;         /* clarity */
   --color-honey: #CE9B4F;       /* voice */
   --color-blossom: #E59ABF;     /* collocation — warm blossom pink */
+  --color-sage: #B7C7B9;        /* mechanics — calm sage, a gentle tidy-up nudge */
   --color-success: #8FCFA8;     /* saved / accepted */
 
   /* Typography */
@@ -234,6 +235,7 @@ button, a, input {
 .petal-suggestion-clarity  { border-bottom-color: var(--color-sky); }
 .petal-suggestion-voice    { border-bottom-color: var(--color-honey); }
 .petal-suggestion-collocation { border-bottom-color: var(--color-blossom); }
+.petal-suggestion-mechanics { border-bottom-color: var(--color-sage); }
 
 @keyframes petal-suggestion-in {
   from { opacity: 0; transform: translateY(4px); }