refactor(scripts): remove test scripts that discovery replaced

[chubes4]

Second slice of #2402 item 5. Retires the package.json test registry that started this whole thread.

#2406 made tests/*.test.{ts,mjs} and scripts/*-smoke.{ts,php} run by convention. That left most test: entries in package.json as dead registry weight — names nothing invokes, wrapping files that now run on their own.

What changes

	before	after
npm scripts	245	81
test: scripts	218	54
package.json	34.8 KB	19.3 KB

164 removed. package.json was the highest-churn file in the repo — 432 of the last 2,815 commits touched it (15%) — precisely because it was the place tests had to be registered. That role is gone.

Retention was computed, not guessed

A script is kept when it is referenced by any of: the smoke manifest, homeboy.json, a workflow, a doc, the README, another npm script, or repository source — or when it does work discovery cannot express, such as a typecheck or a node --test multi-file invocation.

The 54 kept, by reason:

count	reason
26	referenced by a workflow
6	chained by another npm script
5	runs work discovery does not cover
4	homeboy.json
3	typecheck or multi-runner
10	manifest, docs, or source references

Verification

No dangling references. Zero unresolved npm run targets across package.json, homeboy.json, all three workflows, scripts/smoke-manifest.ts, and docs/:

dangling npm-chain references: 0
dangling refs in workflows/homeboy/manifest/docs: 0

Full aggregate green:

npm run check
# [smoke] check: 4 declared, 328 discovered (concurrency 8), 9 serial
# [smoke] check passed: 341 command(s)
# real 9m53.100s

341 commands, same as before the prune — removing the wrappers removed no coverage, which is the entire point.

One extra change

tests/recipe-step-continuation.integration.test.ts moves to the serial phase. It failed once during a verification run under concurrency, then passed 3 of 3 standalone. It boots a runtime and asserts step counts, so contention flips it rather than merely slowing it — same class as the eight already serialized. Catching it here rather than letting it become an intermittent red is worth the line.

Remaining in item 5

Seven excluded tests still need individual fix-versus-delete calls. They fail for unrelated reasons and each needs its own judgement, so they do not belong in a mechanical prune.

---

AI assistance disclosure: authored by Claude (Sonnet 4.5) running in OpenCode, directed by @chubes4. The model computed the retention set across seven reference surfaces rather than pattern-matching names, verified zero dangling targets after removal, and ran the full aggregate twice — the first run caught the contention flake now serialized. All figures are measured. Reviewed by a human before opening.