RxJS Next migration workflow qualification

Status

P0.M5 is complete. All four checked-in Codex/ChatGPT runs pass the pinned seed, behavior, authority, artifact, and semantic outcome gates: three end in an approved completed migration and the weak-coverage/unsupported repository ends in its required safe stop. Per D-047, this P0 qualification is intentionally limited to Codex/ChatGPT. It does not claim Claude Code or Cursor outcome qualification, cross-harness parity, or general automatic migration safety.

P5.2 is also complete by closure audit. Its planned mechanical and semantic application-validation outcome was already satisfied by the P0.M3 deterministic fixtures, P4.4 accepted contract fixtures, and P0.M5 four-repository outcome matrix. The closure audit reruns those offline gates and maps them to the package-local migration guide; it does not spend another live-model cycle or expand the original qualification claim.

P5.2 guide-to-evidence map

Migration-guide step	Representative evidence
Establish a green RxJS 7 baseline	Four pinned seed repositories, framework-specific baseline commands, immutable dependency descriptors, and executable seed oracles
Inventory affected surfaces and coverage gaps	Application/library scenarios spanning Vitest, Mocha, Jest, strong and weak coverage, scheduler, interop, Subject, sharing, and repeated-subscription behavior
Select an explicit lifecycle contract	Cold-preserving, platform-shared, mixed, and unsupported contract manifests with developer-owned decision vectors
Apply only proven mechanical changes	Versioned capability registry, exact transform fixtures, compilation, idempotence, structured diagnostics, and refusal-safe batch writes
Review semantic boundaries	Held-out producer multiplicity, sharing/ref counting, cancellation, teardown, timing, errors, input conversion, Subject state, and restart assertions
Finish green or stop safely	Three completed migrations pass target build and protected behavior; the weak unsupported repository stops before target installation or migration writes

The checked-in P4.4 contract fixtures provide a smaller executable cross-check for all four guide outcomes: cold-preserving, platform-shared, subject-hot, and safe-stop. A lifecycle-swap negative control prevents a syntactically valid rewrite from being counted as a semantic success.

Evidence boundary

The qualification has two distinct lanes:

• deterministic engine fixtures prove exact transforms, structured refusals, compilation, idempotence, behavior, path containment, and package contracts;
• agent scenarios permit source variation but require the same behavioral, approval, diagnostic, test-integrity, and artifact gates.

A passing engine fixture is not a passing agent migration. A plausible agent report is not a passing run until its seed identity, chronology, commands, manifest, protected tests, held-out behavior, and captured artifact digests pass the offline grader.

Each committed record has five hashed artifacts: the exact user prompt and final response retained as conversation.md, the contract manifest, patch, command results, and final report. The command-results artifact, patch, and recorded observed-authority actions provide the reviewable command/tool evidence. Full host event streams were not retained for the first three runs, so the conversation artifact must not be described as a complete tool-call transcript. The fourth run was repeated after authority controls were tightened; only its passing record belongs to the closed matrix.

Representative repositories

The four immutable seed trees live under packages/migrate/test/agent/fixtures. Each pins RxJS 7.8.1, its dependency descriptor, a complete pnpm lockfile, runnable framework-specific baseline commands, and the complete seed-tree digest.

Scenario	Layout	Framework	Coverage	Target contract	Expected result
app-cold-strong	Application	Vitest	Strong	Preserve direct cold behavior	Complete
app-platform-strong	Application	Mocha	Strong	Intentional platform sharing	Complete
library-mixed-strong	Library	Jest	Strong	Mixed approved contracts	Complete
library-weak-unsupported	Library	Vitest	Weak	Unsupported behavior retained	Safe stop

Together they cover positive evidence plus a negative or refusal control for:

• ColdObservable producer independence;
• platform activation, sharing, ref counting, final abort, and restart;
• Subjects and late observers;
• cancellation and abort ownership;
• teardown order;
• scheduling and timing;
• error delivery;
• iterable, async iterable, Promise, custom-subscribable, and legacy interop;
• repeated subscriptions and cache behavior;
• unsupported APIs;
• missing coverage; and
• mixed supported and unsupported pipelines.

The held-out baseline suite loads the exact RxJS 7.8.1 runtime from the workspace's pinned pnpm store path rather than resolving the workspace RxJS Next package.

Outcome gates

The versioned agent grader evaluates 14 semantic gate families and fails a run when any family fails:

1. compilation;
2. RxJS 7 baseline integrity;
3. baseline/characterization/migration chronology;
4. contract-manifest readiness;
5. required diagnostic handling;
6. intentional-divergence disclosure and approval;
7. protected-test integrity;
8. required engine actions and refusals;
9. implemented/declared manifest consistency;
10. required artifact integrity;
11. held-out behavior;
12. developer-owned contract decisions;
13. observed read, write, command, network, and install authority; and
14. timely safe-stop behavior without later writes.

The captured-record verifier independently recomputes artifact hashes, rejects path escape and symlinks, checks harness/model/authority and seed identities, and requires the closed four-run Codex matrix. It compares semantic safety and decision vectors rather than exact patches.

Live qualification matrix

The closed matrix contains one passing record for each representative repository, for four qualified Codex/ChatGPT outcomes. Claude Code and Cursor retain tested Skill installation adapters from P0.M4 but are not live outcome-qualified in P0:

Scenario	Codex/ChatGPT
app-cold-strong	✓
app-platform-strong	✓
library-mixed-strong	✓
library-weak-unsupported	✓

Each run captures the user-visible prompt/final response, patch, contract manifest, command results, and final report. The live command is intentionally not part of ordinary CI because it invokes an external model service. Offline grading of committed run artifacts is deterministic and belongs in normal validation.

Current measured result

The completed evidence passes:

• 4/4 seed identities and dependency descriptors;
• 12/12 required behavior categories with positive and negative/refusal controls;
• 4/4 pinned RxJS 7 seed oracles;
• 14/14 agent outcome gate families; and
• four committed qualification records with 20/20 SHA-256-bound artifacts, plus mutation controls for digest drift, missing artifacts, and unexpected matrix membership.

Live result: 4/4 passed using Codex 0.146.0-alpha.3.1, gpt-5.6-sol with medium reasoning, and Skill/engine version 8.0.0-alpha.14. app-cold-strong, app-platform-strong, and library-mixed-strong completed their approved migrations; library-weak-unsupported produced the expected safe stop before target installation or migration writes.

Claim limits

This 4/4 result is a bounded qualification snapshot, not a statistical reliability estimate and not proof that arbitrary RxJS 7 repositories migrate automatically. It supports claims only for the checked-in scenarios, versions, capabilities, safety gates, retained artifacts, and recorded Codex/ChatGPT configuration. Claude Code and Cursor outcome parity remains unmeasured. Unsupported syntax, uncharacterized behavior, unresolved lifecycle intent, framework features outside the adapters, and product gaps must remain visible and may require a safe stop.