test: name every enforced behaviour in a ledger, and the test that proves it

[ExtraToast]

Closes #67 (parent #21)

What this branch does

docs/requirements.md names every behaviour this repository guarantees, and for each one the test that fails when it stops being true. Ten rows, REQ-001 to REQ-010: the ADR contract, the link contract, the manifest gate, the boundary gate, the pull request title check, the em-dash ban, the test harness guards, the coverage ratchet's integrity, the package-contents gate, and the script-to-workflow wiring check.

A behaviour could previously lose its proof without anything saying so. The coverage ratchet notices a line that stops running, but not a test file renamed, emptied or deleted while the code it covered stayed in place, and nothing tied a sentence in the documentation to the test that makes it true.

• scripts/lint-requirements.ts holds the ledger honest: every row parses, every named file exists and holds at least one test, ids are unique, the stated row count matches the rows, and every id cited anywhere in the tree resolves to a row. Library first, with a one-line CLI guard, the shape the other gates use.
• Each proving test carries a comment citing its id, so an id resolves from the test as well as from the ledger.
• The gate runs in the Contracts job beside the ADR, link and manifest contracts, and in npm run verify. It is a document contract of the same kind as its neighbours, not a code-against-graph check like boundaries. ADR 0102 asks for the script and its job in one pull request, and test/pipeline-wiring.test.ts is what would otherwise have failed.
• This is the companion of the rule ledger in Give every enforced rule an ID, a ledger row and a fixture that proves it fires #29 and deliberately not the same thing: that one lists the rules a tool enforces, this one lists the behaviours a person depends on. The document says so in its opening paragraph.

Decisions

docs/adr/architecture/0103-a-behaviour-ledger-names-what-a-test-proves.md, citing 0102 for the wiring rule.

Two branches came out of scripts/lint-requirements.ts rather than being covered by a contrived test: a ?? "" fallback on a key that comes from the same object being iterated, and an id-shape check already guaranteed by the row pattern, whose capture group is REQ-\d{3}. Neither could fire.

The ratchet rises: statements 97.17 to 97.5, branches 87.5 to 88.79, functions 100, lines 96.88 to 97.26.

Verification

Each acceptance criterion was broken on the rebased tree and restored, not just run:

• Renaming test/adr-contract.test.ts: REQ-001: names 'test/adr-contract.test.ts', which does not exist
• Emptying test/link-contract.test.ts: REQ-002: 'test/link-contract.test.ts' holds no test
• Citing an unknown id: test/link-contract.test.ts cites REQ-999, which no row carries
• Stating nine rows while holding ten: docs/requirements.md: states 9 rows, holds 10

npm run verify clean, 189 tests. actionlint 1.7.7, the version CI pins, clean.

Follow-up

The coverage numbers in README.md now lag these thresholds. #69 owns that file and holds it to the configuration with a test, so the refresh lands there.