Make the tactical scan cap reproducible, and correct what it costs

[idIing]

The shipped docs/known-limits.md concedes that GreedyTactical truncates its
scan above hand size 8, then defends the headline: "Both shop arms share the
layer, so the paired comparison stays valid."

Shared is not symmetric. A large hand is downstream of buying Jokers, so the
stronger arm meets the degraded regime more often. Measuring that changed the
story, and the cap had no CLI surface, so the claim was unverifiable either way.

--score-budget

uv run python scripts/evaluate.py --agent greedy-shop --vs random-shop --score-budget 8000

Defaults to the v1 300, so the benchmark path is byte-identical. Any other
value is stamped jackhammer/tactical-sweep/v1 with scope: diagnostic, never
jackhammer/v1, and the result's agent.tactical records the budget that
actually ran — a sweep cannot be read as a v1 number. Agents the flag cannot
reach (random-legal, or any agent building its own tactical layer) raise
rather than silently no-op.

What it costs

	budget 300	budget 8000
paired delta	+1.567 [+1.400, +1.729]	+1.667 [+1.504, +1.825]

Paired difference-of-differences +0.100 [+0.046, +0.167], excludes zero.
greedy-shop truncates on 576/8341 play scans (6.91%) against 151/5176 (2.92%)
for random-shop. The effect is concentrated: the 201 seeds that never truncate
move by exactly zero with zero variance; the 39 exposed seeds move
+0.615 [+0.308, +0.974].

No published v1 number is restated. +1.567 remains the headline, and
+1.667 falls inside its interval, so no v1 conclusion changes. Raising the cap
stays a v2 question.

The worst case, previously undocumented

The Psychic scores any play of fewer than five cards as zero, and at hand size

= 10 the cap enumerates no five-card subset — so every reachable play scores
zero. It bit 2 of 45 greedy-shop Psychic blinds and 0 of 26 for random-shop:
657P5QGW and PM4RVISW both lost at ante 1 with 0/600, and both clear 720/600
at a raised cap. Visible in the shipped records:

uv run python scripts/inspect_run.py data/bench/greedy-shop.jsonl --seed 657P5QGW

Also corrected

• Severity is a ladder, not a cliff: the runs reach hand size 13, where the
  budget is spent inside k=3 and only high card, pair and three of a kind stay
  reachable. The old text stopped at hand 10.
• The launch's 19/240 and 8/240 reproduce exactly, but came from a terminal-state
  estimate that omitted Troubadour and could not see transient hand size. Stated
  explicitly rather than swapped silently.
• README.md and docs/adding-an-agent.md both promised the paired difference
  isolates the shop policy — "and nothing else", "alone". The evidence makes
  that false, and shipping the admission only in known-limits.md would leave
  the front door overclaiming.

74 tests pass; three new ones cover the override's honesty and its refusals.

[idIing]

Status update, and a correction to the reason this was being held.

This PR was held on the expectation that the fidelity fix would invalidate the numbers it documents to three decimals. It did not. #6 lands protocol v2 — agents are now offered every action the engine makes legal — and all three baselines return byte-identical per-seed outcomes on the 240-seed train split. random-shop and greedy-shop have zero differing summary fields; the headline paired delta is unchanged at +1.567 [+1.400, +1.729]. So this PR's scan-cap numbers still stand.

Independently verified before that: this refactor is behaviourally inert. Over 24 games × 2 arms, main vs this branch with no flag, and this branch with an explicit --score-budget 300, produce raw JSONL identical apart from run_id/started_at. --score-budget is genuinely wired (12 seeds: greedy-shop 3.000 → 3.083 at budget 8000, random-shop unchanged), stamps agent.tactical = GreedyTactical(score_budget=8000), and fails fast on random-legal. The Psychic claim reproduces exactly — 657P5QGW ante 1→3 with three High Card score=0 plays, PM4RVISW ante 1→4 with four — and it is not a sim artifact: bl_psychic … debuff = {h_size_ge = 5} (Balatro/game.lua:280), Balatro/blind.lua:527 returns true below five cards, and Balatro/functions/state_events.lua:614 skips the whole scoring block when debuffed, so the live hand scores zero too.

Two things to fix before this merges:

1. Rebase onto v2. protocol v2: give agents the blind-select and cash-out decisions #6 touches known-limits.md (the scan-cap bullet's closing sentence, which said raising the cap "is a v2 question, not a v1 patch" — now stale phrasing) and renumbers the protocol doc. Expect a small conflict there.
2. One doc command contradicts itself. known-limits.md instructs the reader to run --score-budget 8000, then four paragraphs later to inspect data/bench/greedy-shop.jsonl for the Psychic lock. Both use the default out-dir, so the first command overwrites the artifact the second reads — verified, 3 score=0 lines before, 0 after. Give one of them an explicit --out-dir.