M5 Phase 3: test_api onto the fixture pack, and a docs pass that is not optional

[kjpatel]

Part 4b(ii)b. test_api's failures under deletion go 10 → 4, and three of the four remaining are genuine example tests — their subject is a teaching pack's own wording, so they move with rulepacks/ rather than convert.

_pack() now defaults to the fixture pack and takes a name only for those three.

The find

Writing the money phrasing surfaced a mis-authored guard in the pack I had added one PR earlier:

when: { value: "0.00" }     # matches nothing on a money decision
when: { amount: positive }  # the correct form

A money decision's value is an amount and a currency, not a scalar. The demo did not complain — it fell through to the unguarded case and rendered a plausible answer from the wrong branch.

The fixture keeps the corrected version with a comment naming the mistake, because a pack author will make it and the failure mode reads like success.

Confirmation the edit was inert where it must be: fixtures/receipts is byte-unchanged. Phrasing is presentation and stays out of every hashed body.

One test still needs fixture growth

test_every_placeholder_the_validator_accepts_is_one_the_demo_renders exercises {daysBetween:…} and {fact:…|day}, so the pack needs a date attribute. Cheap, but it is another rebuild — it belongs with part 5's other growth rather than alone.

And the standing policy you asked for

CLAUDE.md's definition of done now requires a documentation pass as the last step of every PR, with a five-question checklist and the reason the ordering matters:

Documentation written before the work is finished describes what you intended; the interesting part is almost always what the work taught, and that is only knowable at the end.

It cites the two cases where that pass caught a claim the code had just falsified — the architecture guide still saying "usually supersedes" after C6 made that unrepresentable, and fixtures/README.md never mentioning a case the corpus had grown two PRs earlier. No need to ask for it again.

Verification

975 passed, 7 skipped
verified 351 cases                        # byte-for-byte
git diff -- golden/ fixtures/receipts     # empty

Deletion re-measured in a detached worktree, which is now the standing method for this phase.

🤖 Generated with Claude Code