Add prose write test iterations

[rawwerks]

Summary

Adds a bounded generated-test loop for apply-enabled prose write CLI runs.

• --test-iterations=1 is the default for prose write --out ... --apply and --run.
• --test-iterations=0 disables the loop.
• --test-iterations=3 allows up to three generated-test attempts with apply-enabled repair authoring passes between failed attempts.

Use Case

When prose write authors a package that includes kind: test files, the CLI should dogfood those tests before handing the result back or starting the --run follow-up. This is the sister PR to #90: #90 adds write/apply/run, and this PR adds the test feedback loop on top.

Design Boundary

This stays a CLI host-adapter macro. --test-iterations is parsed by the CLI and is not passed into prose-author. The CLI invokes ordinary prose test <generated-test> after successful apply-enabled authoring, then invokes a new apply-enabled authoring pass only to repair failed generated tests.

Plain package-only prose write "..." is unchanged. Plain in-session prose write still routes to prose-author; non-CLI hosts that cannot perform an explicit --test-iterations macro must reject it before authoring.

The loop fails closed if a repair pass deletes an initially generated failing test instead of making it pass. Test discovery skips symlinks and tolerates file/directory races while walking generated files.

Repair prompts include captured failing test output capped at 12 KB. The side-effect ban and target-path repair instruction are prepended before captured output to reduce prompt-injection risk.

If the loop exhausts its iteration bound, the CLI returns the final failing prose test exit code rather than a distinct sentinel.

Examples

prose write --out src/reviewer --apply "draft a reviewer system"
prose write --out src/reviewer --apply --test-iterations=0 "draft a reviewer system"
prose write --out src/reviewer --apply --test-iterations=3 "draft a reviewer system"
prose write --out src/reviewer --run "draft a reviewer system"

Testing

• pnpm --filter @openprose/prose-cli test -- tests/prose/command-model.test.ts tests/cli/cli.test.ts tests/skills/open-prose.test.ts passed: 140 tests.
• pnpm --filter @openprose/prose-cli typecheck passed.
• pnpm --filter @openprose/prose-cli build passed.
• git diff --check passed.
• ubs tools/cli/src/commands/base.ts tools/cli/src/prose/command-model.ts tools/cli/tests/prose/command-model.test.ts found no critical issues; remaining warnings were existing heuristic warnings in tests/general code patterns.
• git rebase origin/feature/prose-write-run completed after preserving both the Add prose write run option #90 --out prompt coverage and this PR's test-iteration coverage.

Residual Risk

This PR is stacked on #90 and should be reviewed/merged after or with that branch. The local full CLI suite still has the known tests/prose/crash-window-replay.test.ts timeout observed on #90/main-local testing; the focused slice and CI-equivalent typed/build checks for this feature are green locally. GitHub checks are rerunning on the latest pushed head.

[rawwerks]

Applied the advisory #91 polish on top of the rebase to PR #90 head in 328af05:

• clearer missing-value UX for --test-iterations --apply
• race-tolerant generated test discovery/read paths
• README/body notes that --apply now defaults to one generated-test attempt, repair prompts include capped captured test output, and exhausted iterations return the final failing prose test exit code

Focused tests, typecheck, build, git diff --check, and UBS critical=0 passed locally. GitHub checks are green on the latest head.