Dragon-s-Lair-X68k/docs/STATUS.md

# Status & next-session handoff — end of session 7 (2026-08-23)

## NEXT SESSION: make the mode decision cost-aware

The decoder exists, it is pixel-exact, and **it does not fit**. On the worst
sustained window at the shipping `sasi` profile it costs a mean of **81.7% of a
12fps frame budget** and **31% of frames exceed 100%** (`scsi`: 94.9% median,
42% of frames miss). FINDINGS 28. CPU is now the binding constraint — the first
time in this project that it has been.

The fix is not assembly micro-optimisation. It is that **`vq_hybrid.decide()`
minimises `D + lam*R` — distortion against BYTES — on a machine where the
binding budget is CYCLES**, and the two are not proportional:

| mode | payload bytes | measured cycles | cycles per byte |
|---|---:|---:|---:|
| SKIP | 0 | 13 (clustered) | — |
| V1 | 1 | 300 | 300 |
| V4 | 4 | 448 | 112 |
| RAW | 16 | 400 | 25 |

V4 is **25% of blocks and 50% of the cycles**. The lagrangian charges it 4x a V1
block; the CPU charges it 1.49x. So the encoder currently buys V4 whenever it is
worth 4 bytes, with no idea what it costs to draw.

**The work, in order:**

1. **Add a cycle term to the mode decision.** `decide()` already builds a cost
   matrix of `error + lam * bytes` per mode per block; add `+ mu * cycles`, with
   the cycles vector `[13, 300, 448, 400]` measured in FINDINGS 28.2. One extra
   row of arithmetic in a function that is already vectorised.
2. **Then bisect `mu` per frame against the 833,333-cycle budget**, exactly as
   session 6 bisects `lam` against the byte budget. The machinery is already
   there and already gated: `ratectl.encode_rate_controlled` is frame-driven and
   feeds back the frame it emitted. **But cycles have NO bucket.** Bytes can be
   banked in the ring buffer; a frame that misses its decode deadline is just
   late, because there is no double buffer to decode ahead into. So this is a
   hard per-frame ceiling, not a leaky bucket — simpler than rate control, and
   the two controllers have to run together (raising `mu` moves blocks to SKIP
   and V1, which also *lowers* the bitrate, so the byte controller must see it).
3. **Measure the quality cost.** Everything session 6 did for bytes: what does
   fitting 100% of frames in the CPU budget cost in dB, and does any frame hit a
   cliff? `tools/analysis/11_cpu_budget.py` scores a container without needing
   MAME, so the search loop is cheap; confirm the winner on the 68000 with
   `tools/bench/decode.lua`.
4. **28.5 may not be solvable by the encoder at all.** An all-V1 frame — the
   cheapest possible full redraw — is **110.5%** of the budget. A scene cut
   changes 100% of the screen, so *no* mode assignment fits one at 12fps. Decide
   deliberately: allow one late frame at a cut (the outgoing content is
   unrelated, so it may be invisible), spread a cut over two frame times, or
   drop to 10fps where an all-V1 frame fits. This is a design decision, not a
   measurement, and it needs the user.

**Do not start by hand-optimising `decode.s`.** The hand-derived timings agree
with the measurements to 0.5% on V1 and 1% on RAW (FINDINGS 28.4), so the
inner loop is close to what the instruction set allows; the plausible wins are
single-digit percentages against a 36-point gap. The V4 write pattern is the one
place worth a look afterwards — pairing sub-block rows into `movem.l d0/d2,(a4)`
saves ~16 of 448 cycles.

---

## What session 7 settled

1. **68000 code parses a bitstream and draws frames, pixel-exact.**
   `src/player/decode.s` + `tools/bench/decode.lua`. 120 frames of the Singe
   window decoded in sequence, all four block modes, verified against the new
   reference decoder `tools/encoder/dlx.py`. Because SKIP blocks are claims
   about the previous frame, the last frame is only right if all 120 were.
   In `check.sh` now. **FINDINGS 28.**
2. **It does not fit.** Mean 81.7% of a 12fps frame, p90 116.4%, worst 135.8%;
   31% of frames miss at `sasi`, 42% at `scsi`. Zero-wait-state floor, as ever.
3. **The dual-display-path plan (FINDINGS 24.5/25.6) is withdrawn as incoherent
   — the sixth false premise this project has caught.** The compose path needs a
   RAM copy of the previous reconstruction; the direct path's whole selling
   point is that it keeps none. Mixing them displays stale pixels on **70 of 120
   frames**, worst frame 18.8% of the screen. Every coherent repair is worse
   than not mixing. `tools/analysis/10_pathmix_drift.py`, kept runnable as a
   counterexample and gated in `check.sh`. FINDINGS 28.1.
4. **24.5 also compared a copy against a copy.** Its 53.6% and 76.6% both come
   from `blit.s` and neither includes decoding. Compose = decode-into-RAM *plus*
   the 53.6% blit, so it is strictly dearer than decoding into GVRAM. There was
   never a crossover. The player has **one path and no reference frame**, which
   also gives back 96 KB.
5. **The four block modes cost 300 / 448 / 400 cycles, not one number.** V4 is
   1.49x a V1 block while the mode decision charges it 4x the bytes. The 24.5
   model is 2.03x optimistic at the median. `tools/analysis/11_cpu_budget.py`
   reproduces all four frames timed on the 68000 to within 1 point. FINDINGS 28.2.
6. **The container is big-endian but not aligned, and on a 68000 that is an
   address error, not a slow read.** Frame records are variable-length and laid
   end to end, so their boundaries land on odd addresses. Frame 0 decoded
   perfectly, then the length read for frame 1 vectored into the IPL and sat
   there for 59 emulated seconds looking like an infinite loop. Found by dumping
   PC and the address registers — the code was right, the data layout was not.
   FINDINGS 28.3. **Encoder gap: `encode.py` should pad records to 4 bytes.**
   Measured cost 1.66 B/frame = 20 B/s against 110 KB/s.
7. **A full frame does not fit at 12fps in any mode.** All-V1 is 110.5%, all-V4
   165.2%, all-RAW 147.6%. At most ~88% of the screen can change in one frame
   however cheaply it is coded, and scene cuts change 100%. FINDINGS 28.5.

---

## What session 6 settled

1. **Rate control works, is wired in, and is ON by default.** `encode.py`
   bisects lam per frame under a leaky bucket; `--fixed-lam` restores session 5
   behaviour. FINDINGS 27.
2. **Both overshoots are closed for under 1 dB.** On the Singe window, totals
   including audio: `sasi` 137.4 -> **109.5 KB/s** (target 110) for -0.60 dB,
   `scsi` 381.6 -> **280.0 KB/s** (target 280) for -0.91 dB. Zero frames hit the
   lam=800 cliff at either profile. FINDINGS 27.2.
3. **The FINDINGS 26 desync is gone by construction, not by tuning.** The
   encoder is frame-drivable (`vq_hybrid.frame_ctx` / `decide` / `paint`) and
   rate control feeds back the frame it actually emitted. The regression test
   `tools/analysis/09_ratectl_drift.py` goes 111/120 drifting frames -> **0**,
   and it is now part of `./tools/bench/check.sh`. FINDINGS 27.1.
4. **Rate control makes the display path cheaper.** Raising lam moves blocks to
   SKIP and V1, so there is less to write: `scsi`'s median display cost drops
   53.6% -> 47.1%. The decoder conclusion of 25.6 is unaffected. FINDINGS 27.3.
5. **FINDINGS 26.5 was wrong in both halves, and this is the fifth false premise
   this project has caught.** `_paint` was not the bottleneck (14% of a frame)
   and the ladder was never "minutes" (~18 s; the minutes were k-means in
   `build`). Vectorising it was still right — 17.1x — but what actually makes
   per-frame rate control affordable is that `VQ.assign` depends on neither
   `lam` nor `prev`, so it is cached: a 12-step search over 120 frames costs
   **0.31 s** against 49.1 s. FINDINGS 27.6.
6. **`--prefill` is a trap and defaults to 0.** It buys a permission to overshoot
   of exactly bucket/nframes; at prefill=1.0 the Singe window goes to 116.3 KB/s
   against a 110 ceiling, and on a 14-frame clip it disables the controller
   outright. FINDINGS 27.4.
7. **Fixed-lam `sasi` was already 5% over target on 00020**, the clip everyone
   called easy — nothing noticed because the profile table quotes PSNR, not
   bitrate. FINDINGS 27.5.
8. **1.2-second clips cannot be used to judge rate control.** The bucket's
   startup transient is bucket/nframes: 6% on a 10 s window, 20% on 00020. Same
   lesson as FINDINGS 25.3, different costume.

---

## Start here: is the tree still green?

```
./tools/bench/check.sh
```
~2 min, needs the Blu-ray mounted. Re-runs both display regression tests from
source media **and the rate-control drift test** (session 6), then prints
`ALL GREEN`. Verified green at end of session 6.
If it fails, fix that before doing anything else — everything downstream assumes
the display path is pixel-exact.

## Decisions locked

| decision | value | why |
|---|---|---|
| Target CPU | 68000 @ 10MHz (stock) | hardest honest constraint |
| Display mode | 256 colors, 256x192 in 256x256 CRTC mode | every mode is 1 word-access/pixel, so 256c is free vs 16c |
| Double buffer | **none** — page 1 sacrificed | enables `movem.l` 24px bursts; delta coding needs a RAM reference frame anyway |
| **Codec** | **hybrid VQ: SKIP / V1 4x4 / V4 four-2x2 / RAW, per-block rate-distortion** | flat 4x4 VQ was measured and rejected — see FINDINGS 9-10 |
| **Quality modes** | **two: `sasi` and `scsi`** (USER DECISION, session 2) | one codec, one decoder, one bitstream; only `lam` differs |
| Framerate | 12 fps, **explicit decimation** | source has zero duplicate frames; no free "twos" win |
| Emulator | MAME 0.277 x68000 | accurate enough that measured cycles mean something |
| SNES project reuse | **MIT — cleared** | `data/events/` scene graph is reusable with attribution |

### The SASI/SCSI question is RESOLVED
Session 1 left "which machine do we target" open. The user's answer: **ship both**,
as two quality profiles. This is now implemented rather than hypothetical — the
bitrate ceiling is a build parameter in `tools/encoder/ratectl.py`:

| profile | target | lam | quality (00020 / 00146) | machine |
|---|---|---|---|---|
| `sasi` | 110 KB/s | 60 (floor) | 36.9 / 29.6 dB | stock 10MHz ACE/EXPERT |
| `scsi` | 280 KB/s | 10 (floor) | 39.4 / 32.3 dB | Super/XVI, or CZ-6BS1 board |

**As of session 6 `lam` is a floor, not a setting.** The target is a ceiling and
the encoder bisects lam per frame to stay under it; the profile's lam is the
best quality it is allowed to spend on a quiet frame. On the worst sustained
window that takes `sasi` from 137.4 to 109.5 KB/s and `scsi` from 381.6 to
280.0 KB/s, for -0.60 and -0.91 dB. FINDINGS 27.2.

Sized against the user's working figure of **4 Mbps = 488 KB/s sustained**, on
SD-backed SCSI (BlueSCSI / SCSI2SD) — so that rate is a bus-limited **constant**,
not an average over seek latency.

**Both profiles fit with room.** Ring-buffer simulation on the real per-frame
sizes gives **zero required prefill** for every scene at both profiles: the fill
delivers 40.69 KB per frame time and only one measured frame (42.10 KB) exceeds
that, recovered by the next. A 256 KB buffer carries ~1 s of stall tolerance,
far more than an SD-backed seek needs. FINDINGS 21.

An earlier warning here said `scsi` did not fit because a frame peaked at 96.4%
of the pipe. That compared instantaneous demand to a sustained rate as if they
had to match frame-by-frame; with a buffer the test is cumulative, and it passes.

`scsi` is now within **0.5 dB of the palette ceiling** on 00020. These were
initially set at 45 / 75 KB/s, which was 12% / 7% bus utilisation — read off the
RD curve rather than derived from the hardware. See FINDINGS 17.

Codebooks are **k=256 with 1-byte indices** in both profiles. k=1024 was measured
and rejected — see FINDINGS 14, it was a false-good result from a rate model
that undercharged the index. Do not ship past `lam~800`; FINDINGS 15 has the cliff.

Because of the RAW escape mode, `lam=0` is **pixel-exact** against the palettised
frame (measured 0.00 dB loss). The profiles are two points on one continuous
rate-distortion curve, not two codecs.

---

## What session 3 settled

1. **The display path works and is verified end to end.** First real frame on an
   emulated X68000 screen: `docs/images/x68k_first_frame_compare.png`. Full
   write-up in **FINDINGS 22**. Everything before this session was Python-side
   or a headless `-video none` run, which cannot snapshot at all.
2. **The render is pixel-exact, not merely close.** With monitor contrast at 15,
   all 256 palette entries render exactly as `GGGGGRRRRRBBBBBI` + `pal6bit`
   predicts. That exactness is the regression test — see
   `tools/bench/verify_frame.py`, which exits non-zero if it ever drifts.
3. **Three hardware facts that were previously assumed are now confirmed from
   MAME 0.277 source**, not folklore: the palette word format, the 1024-byte
   GVRAM line stride, and the 256-colour page aliasing in `HARDWARE.md`. All
   three were already written down correctly; they are now cited.
4. ~~**A new quality ceiling was measured** — the 15-bit+I palette alone costs
   38.88 dB.~~ **Superseded by session 4:** that figure assumed the shared LSB
   `I` is always 1. Chosen per entry, the ceiling is **40.81 dB**. FINDINGS 23.3.
5. **Two shell traps that wedged session 2's background jobs** are documented in
   the working-setup section below. They cost ~1.5 h of wall clock and a wedged
   CPU core, and one of them was hit again this session.

## What session 5 settled

1. **68000 code drew a frame, and the blit was measured.** `tools/bench/blit.s`
   + `blit.lua`. The snapshot passes `verify_frame256.py` unchanged — pixel-exact
   in the real 256x256 mode. **FINDINGS 23.5 is closed**: no longer "proven from
   Lua only".
2. **The 38% full-frame blit estimate is dead. It is 53.6%.** And that is a
   zero-wait-state floor — MAME models no GVRAM wait states, so real hardware is
   worse. FINDINGS 24. Every variant was hand-derived from the MC68000 timing
   tables before being measured and the two agree to 0.006-0.43%, so this is not
   another MAME artefact.
3. **Reading the source frame is exactly half the blit cost** (V1 53.6% vs a
   write-only floor V3 of 27.1%). That is what makes the architecture question
   below live.
4. **That number is now measured, and the answer is "implement both paths".**
   On the worst sustained window found on the disc, 30% of frames (`sasi`) to
   53% (`scsi`) sit above the 70% crossover and want the flat blit; the rest
   want direct-to-GVRAM. A player that picks per frame — the mode headers are
   parsed before any pixel is written, so the count is free — pays a **median
   37.0%** and is **capped at 53.6%**. FINDINGS 25.6.
5. **The sustained action sequence exists, was found by measurement, and breaks
   both profiles.** `tools/analysis/07_motion_survey.py` scans a whole stream
   for the hottest sliding window; on 00223 it is t=539.4s, the Singe endgame,
   at 2.01x the stream mean. There, fixed-lam `sasi` overshoots 110 -> 129.6
   KB/s (+18%) and `scsi` 280 -> 373.8 KB/s (+34%). **Rate control is no longer
   insurance — it is required.** FINDINGS 25.3.
6. **The two largest streams on the disc are bonus material, not game footage.**
   00216 is the feature with a burned-in commentary PiP; 00215 is the commentary
   itself. **00223 (9.4 min) is the clean one.** A size-ranked survey would have
   encoded live action. FINDINGS 25.1.
7. **Rate control is unsound as written, caught before wiring it up.** The
   lam-ladder in `ratectl.py` picks frames from independent temporal chains,
   so SKIP blocks reference reconstructions the decoder never saw: 111 of 120
   frames drift, worst frame 43.4%, reported PSNR overstated 0.36 dB. Regression
   test `tools/analysis/09_ratectl_drift.py`. FINDINGS 26.
8. **On hard content the scene palette, not the display, is the binding
   ceiling** — 31.33 dB on the Singe window against 39.90 dB on 00020 and 40.81
   dB for the X68000 display. `scsi` is already within 0.51 dB of it.
   FINDINGS 25.4.

### Superseded within session 5
4a. **The decoder architecture hinged on one unmeasured number.** Writing
   codewords straight into GVRAM costs 76.6% of the frame budget for a *full*
   frame (V4 — the 1024-byte stride kills the `movem.l` burst), but scales with
   the non-SKIP block fraction and needs **no RAM reference frame at all**,
   because the previous frame is already in GVRAM. Compose-then-blit is a flat
   53.6%. **They cross at 70% of blocks changed.** FINDINGS 24.5.

---

## What session 4 settled

1. **A real 256x256 CRTC mode exists and is verified.** `crtc_mode.lua`, derived
   from `x68k_crtc.cpp`'s divisor ladder rather than recalled — the derivation is
   self-checking (368 = 1104/3 exactly, so the horizontal registers divide by
   three with no remainder). Snapshot is native 256x512, active area pixel-exact,
   letterbox true black. FINDINGS 23. The x=512 wrap of FINDINGS 22.5 is gone.
2. **The palette ceiling was wrong by 2 dB, in our favour.** The shared LSB `I`
   must be chosen **per palette entry**, not hardcoded to 1. Doing so lifts the
   display ceiling from 38.85 to **40.81 dB** and is the only way to get true
   black at all (`pal6bit(1) = 4`). 102 of 256 entries want `I = 0`. This
   supersedes FINDINGS 22.4 and gives `scsi` ~2 dB more headroom than believed.
   **The encoder does not do this yet** — see the encoder-gaps list.
3. **Letterboxing costs one palette entry.** 255 colours + a reserved black at
   index 0, with `I = 0` on it. `prep_frame.py --reserve-black`. FINDINGS 23.4.
4. **MAME's graphics double-scan is phase-shifted one raster line** — pairs are
   (1,2),(3,4),..., not (0,1), because `get_gfx_pixel` halves the *absolute*
   scanline and `vbegin = 41` is odd. Cost a false failure. FINDINGS 23.2.

---

## What session 2 settled

1. **The critical-path question is answered.** "Does VQ soften Bluth's linework
   unacceptably?" — **flat 4x4 VQ: yes, badly. The hybrid (SKIP/V1/V4/RAW): no.**
   Verified by eye, not just PSNR. See `docs/FINDINGS.md` 9-11 and the two
   images in `docs/images/`. Both profiles use **k=256**; see item 2b.
2. **Session 1's 12fps bitrate was wrong** (183 KB/s claimed, 340 KB/s measured).
   Halving the framerate does not halve the bitrate. FINDINGS 8.
2b. **A fourth false-good result was produced and caught this session** — k=1024
   codebooks looked like a +2.4 dB free win because the rate model charged 1 byte
   for a 10-bit index. FINDINGS 14. The k=256 configuration ships.
3. **The 256-colour palettised frame is the real quality ceiling** and it looks
   excellent. Judge the codec against that, not against 1080p.
4. Encoder exists and produces a real bitstream: `tools/encoder/`.

---

## Encoder — working

```
python3 tools/encoder/extract.py 00020 /tmp/fr_00020 12 crop
python3 tools/encoder/encode.py  /tmp/fr_00020 out.dlx --profile sasi --preview p.png
```

| file | role |
|---|---|
| `extract.py` | .m2ts -> 256x192 PNGs, 12fps, spatial-only denoise |
| `vq.py` | palette, blockify, hand-rolled k-means (no sklearn on this box), PSNR |
| `vq_hybrid.py` | the codec: 4 block modes + lagrangian mode decision |
| `ratectl.py` | SASI/SCSI profiles, leaky-bucket rate control |
| `encode.py` | CLI + `DLX1` container writer |

`DLX1` container layout is documented in the `encode.py` docstring. All
multi-byte fields are **big-endian** so the 68000 reads them with a plain `move`.

### Known encoder gaps
- ~~Rate control is written but not yet wired into `encode.py`.~~ **DONE,
  session 6.** It is on by default; `--fixed-lam` restores the old behaviour.
  Gated by `tools/analysis/09_ratectl_drift.py`, which is now in `check.sh`.
- **Payload is deliberately NOT entropy-coded** — deflate decode does not fit in
  the 68000's frame budget (FINDINGS 17.2). Do not "optimise" this later.
- **Frame records are not aligned.** They must be padded to a 4-byte boundary:
  unaligned is an ADDRESS ERROR on a 68000, not a slow read (FINDINGS 28.3).
  `prep_dlx.py` repairs it at load time, which a player streaming from disc
  cannot do. The pad is real bytes on disc, so it belongs inside the rate
  controller's accounting. 1.66 B/frame, 20 B/s.
- **The mode decision is blind to CPU cost.** It charges V4 four payload bytes
  and ignores that it costs 1.49x a V1 block to draw. This is the top item at
  the head of this file. FINDINGS 28.2.
- **Palette packing is not implemented in the encoder.** It still emits 24-bit
  palettes; the X68000 word packing happens Lua-side. Whatever writes real
  palette words must pick `I` per entry by minimum squared error (FINDINGS 23.3,
  worth 1.96 dB) and reserve index 0 as black with `I = 0` (FINDINGS 23.4).
- Codebooks are per-scene and rebuilt from scratch; no inter-scene reuse.
- ~~`_paint` is a Python per-block loop.~~ **DONE, session 6** — vectorised,
  17.1x. It was never the bottleneck, though: `VQ.assign` is 78% of a frame and
  `H.build`'s k-means is 51 s of a 55 s run. **That k-means is now the thing to
  attack before the full-disc survey**, not anything in the per-frame path.
  FINDINGS 27.6.

---

## Working setup (unchanged from session 1, re-verified)

**MAME ROMs** — `~/mame/roms/x68000.zip`. Must pass **`-bios ipl10`**.
```
mame x68000 -bios ipl10 -video none -sound none -nothrottle -seconds_to_run 3
```
**Assembler** — `tools/vasm/vasmm68k_mot -Fbin -o out.bin in.s`

**Blu-ray** — `udisksctl loop-setup -r -f DRAGONS_LAIR.iso` -> `/media/reala-misaki/BDROM`
(still mounted as of end of session 2).

**MAME Lua harness** — `tools/bench/*.lua`, working. Three gotchas (retain the
notifier subscription in a global; the stack register is `SP` not `A7`;
`autoboot_script` fires at PC=0 before boot) are documented in FINDINGS.

**Two shell traps, both hit again this session:**
- piping MAME (or any long job) through `grep` block-buffers — write to a file.
- `pkill -f <pattern>` matches your own shell and kills it (exit 144).
  Use `pkill -x` or kill by PID.
- **`pgrep -f <name> | xargs kill` kills your own shell too — exit 144.** Same
  root cause as the `pkill -f` trap above: the shell's own command line contains
  the pattern. **Hit again in session 5**, which makes it four times across three
  sessions. Kill by PID captured at launch (`$!`), or use `pkill -x`.
- **`until ! pgrep -f foo.py; do sleep; done` watcher loops never exit.** The
  watching shell's own command line contains the string `foo.py`, so `pgrep -f`
  matches the watcher itself and the loop spins forever. Session 2 left 11 of
  these wedged for over an hour. Wait on the PID (`while kill -0 $PID`) or on a
  sentinel file the job touches when it finishes -- never on a `-f` name match.
- **`timeout N mame ...` does not kill MAME.** MAME catches SIGTERM and, with an
  autoboot script blocked waiting on a flag that never arrives, never reaches
  its shutdown path. `timeout` without `-k` then waits forever while MAME burns
  a full core at `-nothrottle`. Always `timeout -k 5 N`.

---

## Disk throughput benchmark — still blocked, no longer gating

`IOCS _B_READ` returns -1 uniformly. Full diagnosis and the four untested
hypotheses are in session 1's notes (git history of this file, commit 65112b9);
the ordered plan for retrying is in **`docs/BENCHMARK.md`**.

**Status changed twice this session — read this rather than the git history.**
It was briefly promoted to critical-path while the working bandwidth figure was
misread as 4 MB/s. With the correct figure (**4 Mbps = 488 KB/s**) and the
ring-buffer simulation showing **zero required prefill** for both profiles
(FINDINGS 21), the design no longer hangs on it. Pixel-exact on SCSI is **not**
available at 4 Mbps — it needs 92-97% of the pipe — so there is no longer a
"measure it and maybe ship transparent" decision waiting.

What the benchmark is still worth doing for:
- **Confirming the 4 Mbps figure.** It is user-supplied and its provenance is
  not recorded. Every profile hangs off it.
- **Confirming DMA is actually used.** If transfers fall back to PIO the CPU
  cost rises far above the ~12-15% cycle-steal estimate and CPU becomes the
  binding constraint. This is the worst plausible outcome and the cheapest to
  check — do it first.

**Do not try to get the bandwidth number out of MAME.** Its SCSI/SASI devices are
functional models, not timing-accurate; a KB/s figure from MAME measures the
emulator's scheduler. `docs/BENCHMARK.md` covers the three-tier approach
(MAME validates the path, derivation bounds it, real hardware settles it).

## Display path — VERIFIED (session 3), in a real mode (session 4), by 68000 code (session 5).

The first real frame is on screen: `docs/images/x68k_first_frame_compare.png`.

**Session 5 closed the gap this paragraph used to describe.** GVRAM is now
filled by 68000 instructions and the result is still pixel-exact, and the blit
cost is measured rather than estimated: **53.6% of a 12fps frame**, not 38%
(FINDINGS 24). The remaining caveat is different and narrower: MAME models
**no GVRAM wait states**, so 53.6% is a floor and real hardware is worse.
Full write-up in **FINDINGS 22**. Harness: `tools/bench/show_frame.lua` +
`tools/bench/prep_frame.py`.

Three facts the player MUST honour, none of which were guessable:

| what | where | value |
|---|---|---|
| **Un-hide the graphics layer** | CRTC R20 `$E80028` | clear bit 11 ("G-VRAM set to buffer"); IPL leaves `0x0B16` |
| Colour setup (256c) | CRTC R20 bits 9-8 | `0x0100` |
| **Monitor contrast** | `$E8E001` bits 3-0 | IPL leaves **14**; write **15** or everything renders 7% dark |

The `R20 = 0x0116` value quoted here in session 3 is the **768-wide IPL timing**
with the gate cleared. The shipping value is **`R20 = 0x0110`** — see the mode
table in `tools/bench/crtc_mode.lua`, which is now the single source of truth
for all of R00-R08 and R20.

Bit 11 is the one that cost the most time: GVRAM writes land and read back
correctly while the layer is invisible, so the video controller looks guilty and
is not. Contrast `0` blanks the screen — free fade-to-black for transitions.

Palette format is now **confirmed from MAME source**, not assumed:
`GGGGGRRRRRBBBBBI` (G 15:11, R 10:6, B 5:1, shared LSB I), expanded as
`pal6bit((field<<1)|I)`. With contrast at 15 the render is **pixel-exact**.

Ceiling: the 15-bit+I palette costs **40.81 dB** against the 24-bit palettised
source, once `I` is chosen per entry (FINDINGS 23.3 — session 3's 38.88 dB
assumed `I = 1`). Still the same order as the `scsi` profile's own codec error
(39.4 dB), so `scsi` remains near display-transparent, with ~2 dB more headroom
than session 3 thought.

Snapshot recipe that works (`-video none` CANNOT snapshot):
```
SDL_VIDEODRIVER=dummy mame x68000 -bios ipl10 -video soft -window \
  -sound none -nothrottle -plugins -autoboot_script <script>.lua \
  -snapshot_directory ./snap -snapview native -seconds_to_run 6
```
`-snapview native` drops MAME's LED artwork and gives a clean 768x512 screen.

## Next steps, in priority order

1. ~~**Measure the non-SKIP block fraction.**~~ **DONE, session 5** — FINDINGS
   25.6. Answer: implement **both** display paths and pick per frame; median
   37.0% of the frame budget, capped at 53.6%. Reporting is wired into
   `encode.py`. Original framing kept below because the reasoning still governs
   the decoder's inner loop:
   FINDINGS 24.5: compose-in-RAM-then-blit costs a flat 53.6% of the frame
   budget; decode-direct-to-GVRAM costs 76.6% x (fraction of blocks that are not
   SKIP) and needs no RAM reference frame. **They cross at 70%.** Which side of
   70% the content sits on decides which decoder inner loop to write, so this
   must come before writing one.
   **It needs no new machinery** — the mode decision in `vq_hybrid.py` already
   computes it per frame and simply never reports it. Add the histogram
   (SKIP / V1 / V4 / RAW counts per frame) to `encode.py` output and run it over
   the clips already extracted. Report the *distribution*, not the mean: a
   scene-cut frame is ~100% non-SKIP and a held frame near 0%, and the mean of
   those two is a number describing no actual frame.

1b. ~~**Wire rate control into `encode.py`.**~~ **DONE, session 6.** FINDINGS 27.
   Both overshoots closed for under 1 dB, drift test at zero, `check.sh` gates
   it. The remaining rate-control question is not a defect: whether `--rc-floor
   open` is worth taking on quiet content. It measured as worth **0.00 dB** on
   the Singe window (no frame there is quiet enough to saturate the bucket), so
   it needs a genuinely quiet scene to decide, and it is a quality-per-byte
   judgement rather than a correctness one.

2. ~~**68000 decoder skeleton.**~~ **DONE, session 7.** `src/player/decode.s`,
   pixel-exact over 120 frames, gated in `check.sh`. It answered the question it
   was written to answer, and the answer is no: **it does not fit** — mean 81.7%
   of a 12fps frame, 31% of frames over 100%. FINDINGS 28. The follow-on is
   priority 0 at the top of this file.

2b. **Pad frame records to 4 bytes in `encode.py`.** Not optional: unaligned
   records are an address error on a 68000 (FINDINGS 28.3), and `prep_dlx.py`
   currently repairs it at load time, which the shipping player streaming from
   disc cannot do. The padding is real bytes on disc, so it has to be inside
   the rate controller's accounting, not added after it. 20 B/s at 12fps.

2a. **Re-budget everything against the MEASURED per-mode costs**, not 53.6% and
   not 38%. Session 7 replaced the model twice over (FINDINGS 28.2): the display
   path is not one number times a block fraction, and the median frame is 74.4%
   rather than 36.6%. The original note is kept below because its warning about
   downstream figures derived from a dead estimate is exactly what happened
   again.
   ~~Re-budget everything against 53.6%, not 38%.~~ Several downstream figures
   were derived from the old estimate. The blit alone now eats over half the
   frame at 12fps in the compose-then-blit design, before any decode, and MAME
   models no GVRAM wait states so that is a floor. This may reopen questions
   that were closed against the 38% number — check FINDINGS 17.2's entropy-coding
   rejection, which was argued as "54% LZ4 with no room beside a 38% blit". The
   conclusion gets *stronger*, not weaker, but the arithmetic should be restated.

3. **Full-disc survey.** Now scoped by session 5 rather than open-ended: the
   worst *sustained* window is measured (FINDINGS 25), so what remains is the
   distribution over content, not the worst case.
   - Classify **content / menu / bonus** — not just menu vs content. FINDINGS
     25.1: the two largest streams are bonus material and look like content by
     size, duration and bitrate alike.
   - Run `tools/analysis/07_motion_survey.py` per stream first; it is cheap
     (96x72 greyscale) and gives a hot-window shortlist so the expensive encode
     only runs where it matters.
   - ~~Vectorise `_paint` before this run.~~ Done. The cost to attack now is
     `H.build`'s k-means: 51 s of a 55 s run, and it runs once per scene.
   - ~~Do it after rate control (1b), or it measures an encoder nobody ships.~~
     Rate control is in, so the survey now measures the shipping encoder.

5. **Confirm DMA vs PIO in MAME** (see the benchmark section above) — cheap, and
   the only thing that could still move CPU into the binding position.
6. **Resolve the framing question** (FINDINGS 12: crop vs squash vs wide).
   Needs an eyeball against arcade reference, not a measurement.
7. **Import the scene graph.** SNES project `data/events/` (MIT, cleared),
   cross-checked against DirkSimple (zlib) which transcribed the same data
   independently — diff them to catch transcription errors before committing
   any of it to 68000 tables.
8. **ADPCM audio.** MSM6258, 15.6kHz mono, 7.8 KB/s — already budgeted in
   `ratectl.py`, not yet extracted or encoded.

### Explicitly abandoned — do not re-propose
- ~~Entropy-code the payload.~~ Deflate decode is ~216% of the frame budget on a
  68000; LZ4 is ~54% with no room beside a 38% blit (FINDINGS 17.2). All bitrates
  are raw payload. This also demotes the "247 KB/s lossless" figure in FINDINGS 8
  to a compression upper bound, not a shippable design.
- ~~k=1024 codebooks.~~ False-good result from a rate model that charged 1 byte
  for a 10-bit index (FINDINGS 14). k=256 wins at every matched bitrate.
- ~~Flat 4x4 VQ.~~ Rejected by eye (FINDINGS 9).

## Not yet started
- **A player, as opposed to a decoder.** `src/player/decode.s` parses DLX1,
  dispatches all four block modes and draws pixel-exact frames, but it decodes
  from RAM that Lua pre-loaded. There is no disc streaming, no ring buffer, no
  audio, no timing against the VBL, and no scene branching.
- **Codebook expansion on the 68000.** `prep_dlx.py` does it host-side because
  it is a load-time cost and including it would flatter or damn the inner loop.
  The player must do it: 8 KB + 2 KB per scene.
- ADPCM audio extraction/encoding
- Disk image packaging
- Game logic (scene branching, input windows, death clips)

## Reproducing the 256x256 mode result (session 4)

```
python3 tools/encoder/extract.py 00020 tmp/fr_00020 12 crop
python3 tools/bench/prep_frame.py tmp/fr_00020 tmp/frame256.bin 0 --reserve-black
mkdir -p tmp/snap256 && cd tmp && SDL_VIDEODRIVER=dummy timeout -k 5 90 mame x68000 \
  -bios ipl10 -video soft -window -sound none -nothrottle -plugins \
  -autoboot_script ../tools/bench/show_frame256.lua \
  -snapshot_directory ./snap256 -snapview native -seconds_to_run 6
cd .. && python3 tools/bench/verify_frame256.py
```
Exits non-zero on any drift. Expected: `256x512 native, double-scan exact,
active 256x192 pixel-exact, letterbox true black`, ceiling 40.81 dB.

## Reproducing the display result

```
python3 tools/encoder/extract.py 00020 tmp/fr_00020 12 crop
python3 tools/bench/prep_frame.py tmp/fr_00020 tmp/frame.bin 0
mkdir -p tmp/snap_verify && cd tmp && SDL_VIDEODRIVER=dummy mame x68000 -bios ipl10 \
  -video soft -window -sound none -nothrottle -plugins \
  -autoboot_script ../tools/bench/show_frame.lua \
  -snapshot_directory ./snap_verify -snapview native -seconds_to_run 6
cd .. && python3 tools/bench/verify_frame.py
```
Verified cold from the Blu-ray at end of session 3: exact match, 38.88 dB.
(That 38.88 is correct *for this test*: `show_frame.lua` still packs `I = 1`.
The 40.81 dB ceiling comes from `show_frame256.lua`, which picks `I` per entry.)

`tmp/` is gitignored scratch. The frames are NOT in the repo — regenerate them
with `extract.py`; the earlier ones lived in `/tmp` and do not survive a reboot.

## Reference material on this box (not in the repo)

- **MAME 0.277 source: `~/src/mame-mame0277/`** (tarball `~/src/mame0277.tar.gz`).
  Downloaded this session to settle the graphics-layer question. The files that
  matter are `src/mame/sharp/x68k_v.cpp`, `x68k_crtc.cpp`, `x68k_crtc.h`,
  `x68k.cpp`. **Read these before theorising about X68000 video behaviour** —
  six register-poking attempts failed against a gate that one grep found.
- Blu-ray mounted at `/media/reala-misaki/BDROM` via
  `udisksctl loop-setup -r -f DRAGONS_LAIR.iso`.

---

## Parked ideas (not scheduled, not abandoned)

- **Cliff Hanger, retitled as Lupin III** (user, session 4). Stern's 1983
  laserdisc game was cut from *Castle of Cagliostro* and *Mystery of Mamo* with
  the Lupin branding stripped; a port would restore it. Technically **cheaper
  than this project**: same content class (cel animation, flat colour, hard
  cuts), ~13 min of footage vs Dragon's Lair's ~22, and flatter linework than
  Bluth's, so fewer blocks should escape to V4/RAW. The codec, the display path,
  and `crtc_mode.lua` would all drop straight in.
  **The real cost is media prep, not code:** there is no clean master cut to
  Stern's scene boundaries the way `DRAGONS_LAIR.iso` is, so the footage would
  have to be sourced and cut to match. Not to be started until the CPU path is
  proven — it changes nothing about whether this design works.

## Reproducing the blit measurement (session 5)

```
python3 tools/encoder/extract.py 00020 tmp/fr_00020 12 crop
python3 tools/bench/prep_frame.py tmp/fr_00020 tmp/frame256.bin 0 --reserve-black
tools/vasm/vasmm68k_mot -Fbin -o tmp/blit.bin tools/bench/blit.s
mkdir -p tmp/snap_blit && cd tmp && SDL_VIDEODRIVER=dummy timeout -k 5 900 mame x68000 \
  -bios ipl10 -video soft -window -sound none -nothrottle -plugins \
  -autoboot_script ../tools/bench/blit.lua \
  -snapshot_directory ./snap_blit -snapview native -seconds_to_run 120
```
~25 s wall. Prints cycles/frame and % of a 12fps budget for V1-V4, and snapshots
V1's output. To check that snapshot is still pixel-exact:
`sed 's|snap256|snap_blit|' tools/bench/verify_frame256.py | python3 -`

Not added to `check.sh`: `check.sh` asserts pixel-exactness, and asserting wall
timings there would make the green-light check sensitive to host load.

## Reproducing the decoder result (session 7)

```
python3 tools/encoder/encode.py tmp/fr_singe tmp/rc_fr_singe_sasi_rcprofile.dlx --profile sasi
python3 tools/bench/prep_dlx.py tmp/rc_fr_singe_sasi_rcprofile.dlx
tools/vasm/vasmm68k_mot -Fbin -o tmp/decode.bin src/player/decode.s
mkdir -p tmp/snap_decode && cd tmp && SDL_VIDEODRIVER=dummy timeout -k 5 900 mame x68000 \
  -bios ipl10 -ramsize 2M -video soft -window -sound none -nothrottle -plugins \
  -autoboot_script ../tools/bench/decode.lua \
  -snapshot_directory ./snap_decode -snapview native -seconds_to_run 150
cd .. && python3 tools/bench/verify_decode.py tmp/rc_fr_singe_sasi_rcprofile.dlx
```
~90 s wall. Prints cycles/frame and % of a 12fps budget for four real frames
spanning the non-SKIP distribution, four synthetic single-mode frames, and one
full 120-frame pass; then verifies the last frame is pixel-exact. Expected:
median 73.8%, p90 116.4%, max 135.8%, mean 81.7%; V1 299.9 / V4 448.2 / RAW
400.4 cycles per block.

`-ramsize 2M` matters — MAME defaults to 4M and the locked target is a stock 2MB
machine. `DLX_VERIFY_ONLY=1` drops the timing anchors, which is how `check.sh`
runs it.

Score a container against the measured costs without touching MAME:
```
python3 tools/analysis/11_cpu_budget.py tmp/rc_fr_singe_scsi_rcprofile.dlx
```
And re-demonstrate why there is only one display path (exits non-zero **by
design** — it is the counterexample):
```
python3 tools/analysis/10_pathmix_drift.py                 # 70/120 frames corrupt
python3 tools/analysis/10_pathmix_drift.py --fix direct    # clean, and cheapest
```

## Reproducing the rate-control result (session 6)

```
python3 tools/encoder/extract.py 00223 tmp/fr_singe 12 crop 539.4 10.0
for prof in sasi scsi; do
  python3 tools/encoder/encode.py tmp/fr_singe tmp/rc_$prof.dlx --profile $prof --fixed-lam
  python3 tools/encoder/encode.py tmp/fr_singe tmp/rc_$prof.dlx --profile $prof
done
python3 tools/analysis/09_ratectl_drift.py       # must exit 0, zero drifting frames
```
Expected, totals including the 7.8 KB/s audio allowance: `sasi` 137.4 -> 109.5
KB/s and 27.82 -> 27.22 dB; `scsi` 381.6 -> 280.0 KB/s and 30.81 -> 29.90 dB;
zero frames at the lam=800 cliff in either. ~55 s per encode, nearly all of it
k-means in `H.build`.

The block-mode map now renders the rate-controlled encoder by default:
```
python3 tools/analysis/08_mode_map.py tmp/fr_singe tmp/singe_modes_rc.webm \
        --profile sasi --scale 2                 # add --fixed-lam to compare
```

**Do not judge rate control on `tmp/fr_00020`.** It is 14 frames; the leaky
bucket's startup transient is bucket/nframes, so it lands 18% under target there
for reasons that have nothing to do with the content. FINDINGS 27.5.

## Reproducing the sustained-action result (session 5)

```
python3 tools/analysis/07_motion_survey.py 00223 10        # -> hottest window t=539.4s
python3 tools/encoder/extract.py 00223 tmp/fr_singe 12 crop 539.4 10.0
python3 tools/encoder/encode.py tmp/fr_singe tmp/singe_sasi.dlx --profile sasi
python3 tools/encoder/encode.py tmp/fr_singe tmp/singe_scsi.dlx --profile scsi
python3 tools/analysis/08_mode_map.py tmp/fr_singe tmp/singe_modes.webm \
        --profile sasi --scale 2
```
`extract.py` now takes optional `[start_s] [dur_s]` — needed because 00223 is
9.4 min and the windows that stress the codec are seconds long.

`08_mode_map.py` renders palettised source | decoded | block-mode map at 12fps.
Output format follows the extension; **prefer `.webm`** — GIF re-quantises to
256 colours, which is a poor fit for output whose subject is colour fidelity,
and runs larger. It uses `yuv444p` because the mode map is flat saturated colour
on a 4-pixel grid and chroma subsampling smears exactly those edges.