Reuse decoded native streams and retain precise CI failure evidence

[realmarcin]

Trace decodes each captured stream twice during construction. This change reuses those decoded rows while preserving validation order and public behavior. It also isolates the shared-schema fixture, retains both synthetic native cases after CI teardown, and records the exact metadata associated with the first live-stream acquisition refusal.

The new diagnostic records the selected stream, acquisition stage, offset, opened extent and sampled metadata without further reads. Its optional report field is detached from internal state. Existing refusal conditions, cursor-update timing, ValueError-qualified first-stop text and acceptance limits remain unchanged; diagnostic serialization failure preserves the original refusal.

This targets review/4354-offline-acceptance from adapter 0125eebbc214d0907c3b69d23758cd9441911a78; main does not contain this adapter. No persistent cache, cross-capture state or provider campaign is introduced.

Validation:

• 90 constructor tests passed, including 41 differential cases. The merged comparison in Measure complete Trace constructor parity and decode reduction #4664 verified complete canonical state, public results and refusal graphs; all 184 comparison-driver tests passed. Three alternating pairs measured median constructor thread CPU of 101.3 ms versus 54.8 ms, with first-sample and order/cache limitations retained. This does not measure a whole-path speedup.
• The fixture isolation fix is merged on main in Keep schema hardlink rejection tests inside private fixtures #4669 with 145 related tests; the exact regression also passed on this native base. The measured trace implementation remains unchanged.
• 35 CI-retention controls passed, including actual tiny pytest teardown/exit behavior, partial and missing cases, mode/byte preservation, worker isolation and external-link non-dereference. Retain case root directory mode in native CI diagnostic archives #4677 corrected the missing root-directory archive entry; the initial platform-specific test failure remains documented.
• 105 focused stream/controller tests passed at d01ef23c37e22137be367a73f252cc74d4a58d73, including 26 new reader cases and 15 independent adversarial cases. These verify exact refusal metadata, immutable/detached report data, serialization fallback, unchanged cursor timing, no extra acquisition and refusal of later callbacks. All frozen controller dependency bytes are unchanged. The controls inject metadata schedules; they do not reproduce or diagnose the hosted failure.

Integration remains blocked on required CI and complete native acceptance. The retained run at aa845f2e had two distinct failures: neutral stopped at 253.860 seconds on an equal-size/changed-mtime refusal, while omission reached the existing 900-second attempt deadline after sealing phase 1/core. The neutral artifact does not identify the exact sampled metadata pair or prove a race or rewrite; #4680 remains open. The original archives, matching JUnit, selected source bytes and reviewed standalone verifier are preserved in #4683. Prior deadline failures also remain documented.

Capture and diagnostics do not change test selection, scheduling, helper behavior, admission or 900/180/120-second bounds. Fresh exact-head CI and independent case verification remain required. Scientific and provider-execution obligations are separate.

Closes #4675.
Closes #4677.
Closes #4682.

Refs #4537, #4354, #4400, #4680.