clay-borg/history/260801-cb-wp-0008-log.md

2026-08-01 — CB-WP-0008 delivery notes

Kept out of the workplan file so it stays loadable (loop-lint).

T01 — bots

games/ground/src/bot.rs, 8 tests. Policy + RandomPolicy (seeded ChaChaRng) + GreedyPolicy (stated heuristic, deterministic), a legal_commands generator that filters candidates through validate, and a driver that reaches GameEnded.

Measured, 3-player, seed 42, both policies to GR-R09's fifth round:

policy	commands	events	claimed	total vs threshold
random	36	81	1	2 / 7
greedy	30	51	2	3 / 7

The module is not gated on scenarios; its tests are. A bot needs only the aggregate, but the setup presets live behind that feature. That is a real seam and it is left visible rather than papered over with a feature the bot does not need.

What the second consumer found

INTENT's second-use rule is why T01 was worth doing before the CLI. Three findings, none of them fixed here, because fixing a rule from inside a bot is exactly the invention-in-isolation the rule guards against.

1. GR-A13 admits SOLVE against an already-claimed Problem. The rule names "a face-up, non-Denied Problem"; claimed_by is not mentioned, so check_targeting accepts it and resolution then does nothing. The action is silently wasted. Greedy hit this immediately — it spent all fifteen selections re-solving Problem 1 and scored 1. The policy now avoids it; the rule is untouched and this is a U-item candidate for GroundRules.

2. The 3-player standard fixture cannot succeed. GR-S01 gives 3 Problems at 3–4 players, the preset values them by priority (1, 2, 3 — maximum 6), and GR-E01 sets the 3–4p threshold at 7. group_success is therefore unreachable at 3 players regardless of play. No scenario noticed because no scenario plays to scoring with maximal claiming. The fixture is documented as placeholder ("Problem content is scenario data") — so this is evidence that the placeholder is not neutral, not that GR-E01 is wrong.

3. K13's Project trait has zero implementors. grep finds no impl Project anywhere in the tree. T02 is its first consumer, and the projection will be written against a trait that has never been exercised.

Process

The acceptance tests were mutation-checked by hand (no mutation-check row — that denominator is the GameKernel acceptance table, and adding a game-level row would move a metric by changing its question).

The first mutation was a no-op and SURVIVED: it changed apply(...)? to let _ = apply(...), which still issues the command. Removing the Resolve call outright turned three tests red for their stated reason. Third recorded instance of the weak-mutation class — the retrospective in 260801-instrument-the-table-retrospective.md predicted mutation strength would stay a standing maintenance cost, and this is that cost arriving on the next pass.

T02 — cb-play

tools/cb-play (make play), 6 tests. A human seat is a Policy like any bot, so the CLI adds no second driver: HumanPolicy renders the projection, lists the legal commands, and reads an index or pass.

• K13's first implementor is games/ground/src/view.rs. Hidden: other seats' face-down selections until Reveal, hands (count only), the undealt deck (count only), a face-down Problem's suit and value, and the seed — not secret content, but a seat holding it can compute the deck.
• A played session becomes an artifact. --record FILE writes it as a scenario the runner executes; --replay DIR writes a .cbreplay bundle. games/ground/src/record.rs is the inverse of parse_command, and its warrant is a round-trip test over every command shape — an encoder checked against hand-written expectations only agrees with itself.

Two vacuous tests, caught by counting

The acceptance clause "a seat's projection never contains another seat's hidden selection" passed twice while asserting nothing.

1. The first version asserted the rendered text contained face-down — which it does, from Problems, in every game ever rendered.
2. The counted version then reported 0 inspected entries: seats are asked in order, so a human at P1 is always prompted before anyone has selected and never sees a hidden selection at all.

Seating the human at P3 exercises the rule; the test now counts what it inspected and requires at least four. Mutating the projection (revealed = true) fails it for its stated reason.

This is the same class as the AM-2 expect that matched passing output — an assertion that cannot distinguish the two worlds. The remedy that worked both times was counting what the harness examined, not strengthening the predicate.

T03 — the 2–6 player range

Every scenario in the corpus was 3-player. Now: gr-o01-two-player, gr-o01-six-player, gr-e01-threshold-unreachable-2p, plus all-bot games at every count under both policies, each reproducing at the same seed (K8 at the boundaries, not only in the middle), and the CLI transcript test run at 2p, 3p and 6p.

Nothing broke. The rules are seat-count-generic and the range works. What the boundaries exposed is arithmetic:

seats	Problems	best possible total	GR-E01 threshold	group success
2	2	3	5	unreachable
3	3	6	7	unreachable
4	3	6	7	unreachable
5	4	10	9	reachable
6	4	10	9	reachable

With the standard preset's placeholder values (value = priority), no play at all can clear the threshold below five seats. GR-S01 calls the fixture a stand-in for scenario Problem data, so this is evidence the stand-in is not neutral — not that GR-E01 is wrong. It is pinned three ways: a scenario that passes on the fact, a test that asserts the arithmetic, and a provisional marker with ground-game as owner so it ages in make coverage (now 6 provisional defaults).

The test carries its own delete-by condition: it is expected to fail when Problem values become scenario data, and that failure is the signal to delete it rather than re-tune it.

Two smaller notes, recorded and not acted on: at 2 players GR-L01's second relation slot can never be used (there is only one possible partner), and a human at P1 is always prompted before any other seat has selected, so the seat order makes P1 a weaker test position than P3.