docs: sync for v3.3.0.4 — suite counts, test-helper contract, roadmap

[savvides]

Post-ship doc pass covering v3.3.0.1 through v3.3.0.4. No code changes.

What was stale

Three docs said "eight suites"; there are ten runnable ones now (test-doctor.sh and test-status.sh landed in v3.3.0.4). CLAUDE.md also still claimed smoke-test was 371 assertions — it's 393.

What was undocumented

test/test-helper.sh is the one genuinely new contributor contract in this range and nothing described it. It's sourced, not run, so listing it as an eleventh suite would have been wrong — called out explicitly, along with the rule that a suite must not define its own counters (smoke-test fails on one, and a mutation proves that guard works).

CONTRIBUTING gained a "Writing a new assertion" section: check() now takes an optional expected exit code and output regex, which nothing documented, and the two suites keeping a differently-shaped wrapper (check_version, check_listing) needed saying so nobody adds a third variant.

Narrative

• ROADMAP led with v3.3.0.0 and read present-tense about eight suites in CI. Added the v3.3.0.1–v3.3.0.4 roundup covering the two fixes a user would actually notice, and scoped the CI claim to "eight at v3.3.0.0, ten today".
• TODOS.md now links the three open issues from the PR triage. CONTRIBUTING called TODOS.md "the backlog" while Atomic delete: interrupted idstack-learnings-delete can destroy the whole log #60–Test coverage gaps left by the closed May/June PR batch #62 lived only on GitHub. Pointer, not a copy, so the issue stays authoritative.

CHANGELOG left alone — its "eight suites" line sits inside the v3.3.0.0 entry and was true at that release.

Verification

Every numeric claim re-derived by running the thing, not by reading:

smoke              393/393     integration     48/48
setup               17/17      doctor          13/13
status              22/22      manifest-merge  23/23
version-classifier  27/27      plugin-status    9/9
preamble-python      6/6       mutation       18/18 guarded

CI `run: ./test/` steps: 10        test/*.sh files: 11 (10 suites + test-helper.sh)
gen-skills --dry-run: 22 templates, 0 stale

Also checked every test/, bin/, templates/, docs/ path referenced in the docs resolves to a real file.

VERSION unchanged at 3.3.0.4 — docs-only, nothing here to ship.

Note: the usual cross-model doc review didn't run. codex is installed but its vendored binary is missing (ENOENT on codex-darwin-arm64/.../codex), so it fails to spawn. Worth a npm i -g @openai/codex reinstall before the next review-dependent run.

🤖 Generated with Claude Code