protocol v2.1: re-pin the engine, and make preview_play the engine again - #11

Closes #9.

Two coupled changes that each move published numbers, landed together so the battery is
re-baselined once rather than twice, and shipped as protocol v2.1 — a point release: the
contract (items 1–7) is untouched, only the conditions under it moved.

1. Engine re-pin: 4d6f19d → de733eb

The fork is rebased onto canonical Jackdaw 8712c1e, taking eleven upstream commits (including the
O(n²) get_x_same hoist and stake-sticker flag handling) and dropping two of its own that upstream
had absorbed.

The pin on its own moves nothing. Run with the new engine and the scorer untouched, the full
240-seed battery returns greedy-shop 3.204, random-shop 1.637, paired
+1.567 [+1.400, +1.729], 15,349 decisions — the v2.0 headline to the digit. Everything below is
therefore the scorer fix.

2. preview_play builds the scorer's game_state the way the engine does (#9)

GreedyTactical ranks every candidate play through preview_play, so this is inside the reference
agent's play selection, not a diagnostic path. #9 reported three keys score_hand reads that
preview_play never set. Auditing the class rather than the instance — intercepting score_hand on
both paths at real positions and diffing the two game_state dicts over the nineteen keys the
scorer actually reads — found four defects, at 99 of 99 audited positions:

The fourth is not in #9. The engine increments both hand counters in the event queued after
evaluate_play (game.py:704-708), so every scoring context reads pre-increment values;
preview_play passed the post-increment one, which is exactly the value DNA and Sixth Sense test
against zero.

Measured effect before the fix, over six battery seeds: 218 of 36,941 previewed plays disagreed
with the engine's own dry-run of the same play — all 218 a Loyalty Card ×4 the engine applied and
the preview did not. After: 0 of 36,941, and 0 key disagreements.

The re-baselined battery

Full 240-seed train split, attributable (engine.commit de733eb, dirty false, no
"not attributable" warning), stamped jackhammer/v2.1:

Only greedy-shop moves. random-legal never runs the tactical layer; random-shop does, but its
recorded end-of-run inventories never hold any of the five jokers concerned (greedy-shop's do in
23 of 240 runs), and its per-seed outcomes and whole decision histogram are unchanged.

The regression gate

tests/test_exact_score.py, two forms because they fail independently:

• class gate — diffs the engine's game_state against the preview's, key for key, at real
  positions. Catches a missing key and a present-but-wrong one.
• outcome gate — preview total and hand type vs the engine's dry-run, over every play the
  tactical enumerates, on a window pinned to the board that actually diverged.

Both are red on the unfixed scorer and green after; 3.6s.

Everything in docs/known-limits.md was re-measured, not assumed

Every figure in that file was re-derived, not assumed. Almost all of them come back unchanged,
which is the useful result — the scan-cap audit is independent of the scorer:

• paired shifts at --score-budget 8000: greedy-shop +0.104 [+0.046, +0.175], random-shop
+0.004 [+0.000, +0.013]; difference-of-differences +0.100 [+0.046, +0.167]; 13/240 vs 1/240
  outcomes changed
• 201 seeds where the cap never binds → dod exactly zero at zero variance; 39 exposed →
  +0.615 [+0.308, +0.974]
• truncation: 576 scans in 30/240 games for greedy-shop, 151 in 14/240 for random-shop; max hand
  size 13; 0 truncation at 8000
• Psychic lockups: 2 of 45 greedy-shop Psychic blinds, 0 of 26 for random-shop; 657P5QGW and
  PM4RVISW still 0/600 at ante 1 and still clear 720/600 at the raised cap
• arms bit-identical through the ante-1 Small blind in 240/240; first divergence at a shop decision
  in 232/240, never a tactical one

What moved is only what the scorer touches: greedy's scan count 8341 → 8311 (6.91% → 6.93%
truncating), its decision histogram, the raised-cap plateau 3.308 → 3.300, and the sweep delta
+1.667 → +1.658. The 300→8000 cost was re-timed on this machine: 9.6% more combos for 9.6% more
wall clock (63.6s → 69.7s, 14 workers).

Why a point release and not v3

CONTRIBUTING.md requires a new protocol version for an engine-pin or evaluation-procedure change,
and docs/protocol-v2.md forbade editing a numbered item in place. Both are satisfied by v2.1, and
the distinction is now written into the protocol and into provenance.PROTOCOL: a major bump
means a clause changed meaning, so the two versions answer different questions; a point bump
means the contract is untouched and the conditions under it moved, so every figure is re-baselined
while still answering the same question. Without that rule a routine pin move would ship v3, v4, v5.

Checks

uv sync --locked                       clean
uv run ruff check src scripts tests    All checks passed
uv run ruff format --check ...         28 files already formatted
uv run pytest                          98 passed
scripts/evaluate.py --limit 1 -w 1     exit 0  (the CI smoke)
full battery, train, 240 seeds         3 agents, 0 fails, 0 fallbacks,
                                       protocol jackhammer/v2.1,
                                       engine de733eb source=installed-vcs dirty=false,
                                       kit 1.1.0, no attributability warning

The headline reproduced identically across four independent full-battery runs.

Note for the merge

pyproject.toml, CITATION.cff and the README install snippet move to 1.1.0; the v1.1.0 tag
has to be created at merge for that snippet to resolve.