SLM-267: streaming uniform-policy ProgramSpec corpus CLI + saturation reading

[Tyler-R-Kendrick]

Summary

First increment toward VSD2-02's 10k/100k/1M compiler-inverted ProgramSpec corpus matrix (SLM-267).

• scripts/generate_programspec_corpus.py — a resumable, streaming CLI over the existing typed ProgramGenerator: deterministic seeded generation, exact disk-backed dedup keyed on ProgramSpec.id, an immutable rejection ledger, resumable state.json, a content-addressed manifest.json, all under the canonical DataStore programspec root. Uniform policy only.
• tests/test_scripts/test_generate_programspec_corpus.py — 5 tests covering dry-run, dedup/determinism, monotonic resume, config-mismatch rejection, and saturation detection.
• docs/design/compiler-inverted-program-data.md + docs/design/iter-slm267-uniform-saturation-20260725.{json,md} — the iron-law design doc pair for this run, with an honest inconclusive disposition for hypothesis VSD-H7a.
• src/slm_training/resources/versions.json — registers new component data.progspec.corpus_stream v1.

Findings

A measured local smoke run (two capped 2.5-minute shards) found:

• The current default candidate grid is only 1,781 roots and is seed-independent — dominated by pairwise/prop-target coverage, not by depth/width settings.
• The MVP resume mechanism is replay-based rather than cursor-based: shard 2 spent its entire budget replaying shard 1's calls and produced zero net-new records.

Neither the 10k target nor the 1,781-candidate grid was exhausted, so VSD-H7a is disposed inconclusive rather than claimed as a win. Both blockers (small default grid, replay-based resume) are documented as concrete next-increment work in docs/design/compiler-inverted-program-data.md. No gate was weakened and no ship/checkpoint readiness claim is made.

Test plan

• pytest tests/test_scripts/test_generate_programspec_corpus.py — 5/5 pass
• python -m scripts.verify_version_stamps --check --base d3d0a4f — version-stamp gate passes (1 component bumped)
• Manual dry-run + two capped local shards of scripts/generate_programspec_corpus.py (results captured in docs/design/iter-slm267-uniform-saturation-20260725.json)

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
slm-training	Ready	Preview, Comment	Jul 27, 2026 1:55am

[coderabbitai]

Warning

Review limit reached

@Tyler-R-Kendrick, you've reached your PR review limit, so we couldn't start this review.

Next review available in: 8 minutes

Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available.
You're only billed for reviews past your plan's rate limits ($0.25/file).

How can I continue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews.

How do review limits work?

CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability.

For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window.

Please refer docs for additional details.

Review details

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 3f4bd183-c4a5-4d6e-8f70-67ee40476830

📥 Commits

Reviewing files that changed from the base of the PR and between c7f6aaf and 6bd8f7a.

📒 Files selected for processing (6)

• docs/design/compiler-inverted-program-data.md
• docs/design/iter-slm267-uniform-saturation-20260725.json
• docs/design/iter-slm267-uniform-saturation-20260725.md
• scripts/generate_programspec_corpus.py
• src/slm_training/resources/versions.json
• tests/test_scripts/test_generate_programspec_corpus.py

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch claude/great-dirac-v82ph9

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[Tyler-R-Kendrick]

CI red on Vercel deploy check — this diff doesn't touch the dashboard app (src/apps/dashboard), and the identical failure reproduces on #891 (unrelated files), so it's a pre-existing/base-branch-wide infra issue, not caused by this PR. Will re-run when it recovers. The CodeRabbit comment above is just a rate-limit notice (no review content), no action needed there.

---

Generated by Claude Code

[Tyler-R-Kendrick]

Sound work, but not mergeable as-is — needs a rebase with a real (code-semantics) conflict resolution.

Verification at head acdfc8b: PR's own tests pass (5/5, tests/test_scripts/test_generate_programspec_corpus.py), verify_version_stamps ok, repo_policy ok. Docs are honest: inconclusive disposition, no checkpoint/ship claims, blockers filed.

Blocker: PR #893 (merged, 0cb3718) landed a different scripts/generate_programspec_corpus.py and tests/test_scripts/test_generate_programspec_corpus.py at the exact same paths — a 141-line wiring-evidence fixture CLI (--mode plan-only/fixture, Slm267ProgramspecCoverageScalingManifestV1, claim_class=wiring). This PR's streaming/resumable CLI is an add/add conflict with it: both define main() with incompatible CLI surfaces. Merging would either revert #893's fixture or lose this PR's core work — that is a code-semantics decision for the author, not a mechanical versions.json fix.

Suggested resolution: rebase onto main and either (a) rename this CLI (e.g. generate_programspec_corpus_stream.py) so both coexist, or (b) absorb #893's plan-only/fixture modes into this CLI as subcommands and keep its manifest schema + tests intact. versions.json: keep union of components (data.progspec.corpus_stream new v1 alongside whatever main has since added). Happy to re-review immediately after the rebase — the substantive content here (streaming/resumable uniform-policy CLI, honest saturation reading of VSD-H7a) looks good and is not superseded by #893.

[Tyler-R-Kendrick]

Rebased onto main (now at d3965f8) and resolved via option (b): scripts/generate_programspec_corpus.py is now dispatched by --mode, defaulting to stream (this PR's resumable/dedup CLI, unchanged behavior/flags), with plan-only and fixture absorbed from #893 as-is (identical output schema, same Slm267ProgramspecCoverageScalingManifestV1/campaign payloads, same design-doc side effects). Both test suites now live in one merged file (7 → 19 total tests after also merging the harness's own suite) and all pass unmodified in behavior.

versions.json: data.progspec.corpus_stream (v1, unreleased — this PR's first landing) now owns the merged script + test; harness.experiments.slm267_programspec_coverage_scaling bumped to v2, dropping those two now-shared paths and keeping just its harness module, its own test, and its design docs — no more dual-claim.

Verified at new head 292118c:

• pytest tests/test_scripts/test_generate_programspec_corpus.py tests/test_harnesses/experiments/test_slm267_programspec_coverage_scaling.py — 19/19 pass
• python -m scripts.verify_version_stamps --check --base origin/main — ok
• python -m scripts.repo_policy — ok

Ready for re-review.

---

Generated by Claude Code