[rockyshi1993]

Scope

Evidence-led first-pass audit, not a final conformance sign-off.

• Recover and independently hash-verify two confirmed source files: Requirements V0.3 (56,883 bytes; SHA-256 3a0adc3edb41869a6309bae29ec5496df00970c815f8dc16330a90cbf26e703e) and Technical Design V0.6-R5 (114,621 bytes; SHA-256 09f3a5698e238d10b80efb8e26d1c41e51b663ca620efa70815dd7b52f2106ba).
• Parse and preserve every source T01–T34, R01–R22, S01–S14 row; map T cases to V0.6 §15.1, current source/test entry points, honest P/scripted vs N/native evidence tiers and approved P0-B/P1/P2 deferrals.
• Index V0.6-R5 §§1–18 without claiming that document coverage proves implementation.
• Include crosswalk of the previously approved P0-A Implementation Plan V0.1 and today's implemented M1–M5: don't restart planning/reimplement already merged work.
• Explicitly track native Grok/Trusted Host/Windows OS identity security as still BLOCKED.

First-pass findings requiring code-and-contract review

• CF-01 / P1 P0-A: Confirmed Technical Design §4, §16 and approved A-M5-02 require adminTerminate and a separately trusted admin tombstone controller. No such method appears in current src/interactive/index.mjs, coordinator factory result or types/interactive.d.ts. Do not count M5/S12-A as fully PASS.

• CF-02 / P1 candidate: Confirmed Requirements §7.1 and Technical Design §4 specify importable start/complete/resume; current ./interactive exports factory APIs instead. Need check whether a later, user-approved contract change exists before proposing a correction. Do not silently revise the official sources.

• CF-03 / N-BLOCKED: Scripted Host tests and earlier live-Grok file exchange do not demonstrate native Grok JS tool-loop provenance or true Windows security isolation.

• CF-04: Action/Human/Loop/Parallel belong to future approved stages; do not flag their absence as a P0-A defect.

• CF-05 / P1 candidate: Confirmed V0.6 §4/§5 says model/public SDK frames cannot contain complete State and trusted terminal result needs separate auth. Current coordinator asPublicFrame() embeds {state:cp.state} in terminal frames; this is a contract/potential misdisclosure risk, not evidence that the model has actually seen private State. Host-only getPublicAgentTask() remains guarded.

• CF-06 / P1 candidate: Confirmed V0.6 §5/§8 defines an asynchronous five-operation CheckpointStoreV1; current types/interactive.d.ts and Coordinator require synchronous createRun/putArtifact/getArtifact/load/compareAndSwap, directly consume return values, and therefore do not accept an asynchronous compliant Host store. This matters for future host-selected remote persistence. Check whether an explicitly approved contract change exists before any Core adaptation.

Safety and merge gates

Docs only. No production Core, Program ABI, Store schema, public API, frozen requirements or technical design changes. This PR must remain Draft until exact-HEAD CI and independent review verify source traceability and initial findings; even a green CI does not close Phase 2 formal conformance or M6/C12/S13. Do not release, deploy or start a production code fix from this audit.

R2 focused source re-review — 2026-11-12

• New docs-only R2: interactive-phase2-conformance-audit-r2-20261112.md, commit 9ae5efa16f8c71527a1ca4084c1c7819bd9434f5.
• Confirmed the five critical source blobs are byte-identical between R1's older base ae5fc5a and current main 13d573c; the audit baseline is updated in R2, not rewritten in R1.
• Re-scoped CF-02 as an API-shape candidate (V0.3 explicitly says shown signatures are illustrative; functionality exists on the factory); CF-01/05/06 remain implementation/contract differences requiring the original V0.6-R5 and approval record comparison.
• Added CF-07 (P1 candidate): coordinator.start snapshots caller-owned state and startIntent after awaiting Program load, contrary to the R04 first-await snapshot invariant; source-based inference, failing regression not yet executed.
• Added CF-08 (P2 candidate): compiler checks snapshotOptions.featureStage then after awaits passes live options.featureStage to lowering; can cause input-mutation-dependent false rejection, not a proven P0-B bypass.
• Current docs-only exact-HEAD CI: run 43151709088 (check actual result before merge).
• This PR remains a partial Phase 2 audit. No independent re-review of R2, no complete per-case T/R/S execution proof, and no Windows/Grok native security evidence. Do not merge merely because CI becomes green; no production API/schema/semantics change is authorized by it.

---

Note

Low Risk
Documentation-only PR that records audit findings and review gates; it does not alter runtime behavior, APIs, or schemas.

Overview
This PR adds two docs-only audit artifacts—no production Core, types, tests, or schema changes—and explicitly states the work is a partial evidence pass, not a conformance PASS.

R1 introduces a baseline matrix against hash-verified Requirements V0.3 and Technical Design V0.6-R5: indexed T01–T34, R01–R22, and S01–S14 with code/test entry points, honest P-PARTIAL / DEFERRED / BLOCKED statuses, and priority findings CF-01–CF-06 (e.g. missing adminTerminate, SDK export shape vs contract, terminal state on public frames, sync vs async Store).

R2 rebaselines on current main, confirms five critical source blobs unchanged since R1, refines CF-02 (factory exposes start/complete/resume; dispute is top-level export shape), and adds CF-07 (coordinator.start snapshots state/startIntent after the first await) and CF-08 (compileInteractive may use live options.featureStage after awaits). It links PR #17 as intended-failing R04 snapshot tests and records next review gates without authorizing merge or Core fixes.

^{Reviewed by Cursor Bugbot for commit 90cc0de. Bugbot is set up for automated code reviews on this repo. Configure here.}