[EHotwagner]

Add LEARN-02 rubric/routing v1, a dispatch-recording tool, retrospective calibration, and a framework Atlas record. Classification freezes the owning Delivery body before dispatch, preserves unknowns and safety floors, and records the chosen runtime/model/profile and propensity. Recording prepares a digest-bound native comment and existing-schema telemetry observation; explicit live mode uses shared durable custody, native readback and no automatic uncertain-send retry. No learned routing or exploration activates itself.

Calibration uses 7 recorded PR examples from 6 original Deliveries (19 review rounds), plus 32 eligible Deliveries from the attributed-store population. Review correlation is 0.2283, or 0.0509 after one mean per Delivery; partial log-token correlation is 0.2177 across 32 observations, versus −0.8173 in the 4 anchor overlaps. Complete blocking counts and whole-Delivery costs remain Unknown. The report recommends prospective frozen observations rather than claiming policy benefit. Raw evidence remains private.

Final source is 00e2f18820c166997b943f445486a10eef9c96b1 on base b8a444ee9ae7bef535adfcb39445577bcf0feb4e, including coherent 0.114.4. Independent source and assembled-candidate review by or_5245 passed. All 524 base Atlas records remain, plus the two reviewed LEARN versions; current managed manifests and prior writer-roster evidence are preserved.

Validation on the exact candidate:

• LEARN offline suite: Ran 20 tests / OK.
• Original full semantic suite: routine-delivery board mutate: 73/73 mutants killed; skill-quality fixture: all 60 rejection cases passed. Local exit 0 in 395.560143 seconds under enforced 1 CPU / 2 GiB / swap 0 and a 1200-second outer bound, using SDK 10.0.403 / .NET 10.0.14 / Python 3.14.8 with private ext4 outputs. Its three canonical source hashes equal the original Owner briefs carry the reviewers' standard adversarial checks; dispatch names shared-type owners #5239 qualification; no skill-quality / semantic-skill-quality times out at its 10-minute limit (≈40% of runs cancelled, incl. main) #5228 optimization was used.
• Original hosted semantic run 41445223497 SUCCESS, printing the same 73/73 and 60 summaries. Edge run 41445224097 SUCCESS: 66/66 producer-fence tests and 961/961 edge tests.
• All 10 native required contexts pass. Exact rebound eligibility 41445249227 completed after the preserved old-marker failure 41445216920.
• Genuine projection/driver-manifest check, 75-source writer census, 25-entry external availability join and 29/29 skill-root parity pass. The recorder’s live/legacy-route availability remains explicitly not-exercised / gs2089: unresolved; no operational authorization is inferred.

Fixture clock/source/sandbox: synthetic snapshots use fixed 2026-01-01T00:00:00Z and example/repo#1; native/receiver IO is mocked; mutation and CLR fixtures use disposable private directories and original inputs. No credentials or raw telemetry are committed. Earlier failed/cancelled attempts remain retained and linked in the PR history; no hosted check conclusion, gate or timeout was weakened.

This merge supplies the source slice. Root will activate real recording on the next genuinely new dispatch after merge, preserving its actual input-time boundary and checking both comment and telemetry readback; existing assignments will not be backdated. Full Delivery acceptance remains open until that intended-use assertion is observed. Parent #5248 retains prospective learning, policy evaluation and broader adoption.

Refs #5249
Refs #5248
Related semantic-runtime work: #5228

[EHotwagner]

Independent source review PASS by or_5245 for exact candidate 1e20ad9aee5ff1111daa42e308f7da5c95e0789b against base 0f080b8564434bd33f55ea3da0aa6fd176693319: no blocking defects found.

The reviewer verified all three repairs: shared durable possible-send custody and native comment readback; reconstruction of the entire closed telemetry packet before effects; and the corrected attribution join with deduplication and conflicting-assignment refusal.

Recounted evidence: 66 private evidence digests, 0 mismatches; SQL controls 74 lineage rows / 74 distinct assignments / 0 ambiguous invocations / 34 attributed identities, with retained aggregate-row equality. Calibration recount: 6 original Deliveries / 7 anchored PR examples / 19 recorded rounds / 4 token overlaps, plus the broader 32-Delivery cohort. The reviewer checked 20 test cases and the exact-head Ran 20 tests / OK log; Atlas has 526 records and 6 new immutable citations. No private raw evidence is included here.

No edits, tests, or live effects were performed by the reviewer. Hosted required checks remain pending; this PR stays draft until they pass. Root owns merge and acceptance. No live dispatch-recording effect is authorized by this review.

[EHotwagner]

Independent source and final assembly review: APPROVE SOURCE, reviewer or_5245, exact 00e2f18820c166997b943f445486a10eef9c96b1 against b8a444ee9ae7bef535adfcb39445577bcf0feb4e. All eight LEARN patches are preserved; 19 paths retain all 524 base Atlas records plus two reviewed versions, current-base census/manifest entries and previous roster repairs. No blocking findings.

Exact candidate original full semantic qualification passed locally (exit0,395.560143s) and hosted in 41445223497: routine-delivery board mutate: 73/73 mutants killed; skill-quality fixture: all 60 rejection cases passed. Hosted edge 41445224097 passed with66/66 producer-fence and961/961 edge summaries. All ten required native contexts passed. Prior failures remain historical; no timeout or test weakening.

Merge accepts source only and retains delivery-continues. #5249 remains open for the first genuine post-merge dispatch recording and both native comment/telemetry readbacks. No mock or retrospectively backdated assignment establishes that operational effect.