J3 D8-d: T4 caster wall diagnostic — HP+AC, not transit/heals

Plan update only. Diagnostic in sim_results/d8d_findings.md
(gitignored). n=3-5 per caster L10 underdark:

- Multi-region transit eliminated: all casters TPK in r1
  before any cross-region transit; Combats=0 (no elite/boss
  reached).
- Heal-stock not the lever: bumped simConsumableBundle T4
  2 -> 5 Spirit Tonics, +3-5 rooms median, 0/5 clears.
  Reverted.
- Confirmed lever: computeAC class floors (casters 11-14
  vs martial 16-17) + d6/d8 HP scaling.

Recommended next session: lift floors in computeAC
(dnd.go:195) — cleric/druid 3->5, bard/warlock 1->3,
mage/sorcerer 0->2. Validate vs d8prereq_corpus.

gogobee_long_expedition_plan.md

@@ -186,7 +186,7 @@ Caster/martial gap closed from ~40pp to ~22pp. T2/T3 zones are now the picker pa
 - Subtle gotcha: casting a new concentration spell drops the old one. Picker is stateless (one decision per turn) so it can't easily know "I'm already concentrating on X." Acceptable v1: let it overwrite — sim noise will eat the rare double-cast. Stateful version: read `sess.Statuses` for an active concentration marker; defer.
 - Validation: re-baseline against `sim_results/d8prereq_corpus.jsonl`, **not** d7d. Expected: incremental movement at T3 (already high after prereq); unclear lift at T4 (depends on whether picker-gap is the binding constraint there — see D8-d).
 
-**D8-d (TODO next session): T4 caster wall diagnostic.** Every caster still clears underdark 0% post-prereq. Picker is no longer the suspect. Pick cleric (largest prereq gain, still 0% T4) + one underdark run with `-trace`. Hypotheses to disambiguate: (a) multi-region transit damage on caster HP pools, (b) T4 monster attack-bonus vs caster AC 12–14, (c) HP scaling — d6/d8 hit dice can't absorb a T4 damage roll, (d) heal-consumable threshold too low / not enough heals stocked at L10. Likely a one-session triage that lands a tuning knob (consumable stock or AC floor), not a code rewrite.
+**D8-d (DIAGNOSTIC SHIPPED 2026-05-28):** T4 caster wall diagnosed; no code change landed. Writeup: `sim_results/d8d_findings.md`. **Hypothesis disambiguation across cleric/bard/mage/sorcerer/warlock/druid L10 underdark (n=3–5 each):** (a) multi-region transit damage — REJECTED. All casters TPK in r1 (`underdark_surface_tunnels`) before any cross-region transit; `Combats=0` (the elite/boss session table is empty); deaths are entirely inline `SimulateCombat` mob rooms. (b) T4 attack-bonus vs caster AC — confirmed contributor. `computeAC` floor gives casters AC 11–14 vs martial 16–17; T4 standard roster (Atk +5 to +7) hits ~60–65%. (c) HP scaling — confirmed contributor. Caster HPMax 93–110 vs martial 141 (~30% gap). (d) Heal-stock — minor lever, NOT the lever. Direct experiment: bumped `simConsumableBundle` T4 from 2 → 5 Spirit Tonics, re-ran n=5 per class. Median rooms-before-TPK lifted +3–5 but 0/5 clears across every caster. Reverted. **Recommended tuning lever (next session, NOT shipped here):** lift `computeAC` class floors (cleric/druid 3 → 5, bard/warlock 1 → 3, mage/sorcerer 0 → 2) — single function, easy to revert, doesn't move the martial leaders. Validate against `d8prereq_corpus.jsonl`. Hit-dice rescale is the second option but is bigger surface area.
 
 **D8-e (TODO next session): martial T4/T5 regression triage.** Paladin -18pp T4 / fighter -6pp T4 -5pp T5 / ranger -13pp T5 — the cost of moving boss/elite fights from inline `SimulateCombat` to the turn-based `!fight` engine. Two possibilities: (1) turn-engine plays martials slightly worse than inline-sim (action economy / multiattack handling differs — see [[project_sim_multiattack_gap]]); (2) inline-sim was over-rosy and the new numbers are the real ones. Read: trace one paladin underdark run pre/post and diff damage exchanges. If turn-engine is correct, prod martial numbers were inflated and we should not "fix" the regression.