Make finite runtime log diagnosis reproducible and explicit about gaps #4349

🤖 Generated by the Agentic Engineer

Problem and impact

Part of #4329. Repeated read-only runtime investigations currently depend on an ad hoc parser. It can omit unmatched or ambiguous operations, making a finite log sample appear more complete than it is. Operators need reproducible timing evidence before choosing a repair; runtime timings alone do not identify the requests behind an observability incident.

Smallest useful change

Add an explicitly enabled, offline diagnostic for finite runtime-log files. Pair operation starts and results conservatively, report retained time bounds and every parse, orphan, failure and ambiguity gap, and keep input/output private. It must never fetch credentials, change running resources, or declare an incident resolved.

Acceptance criteria

• Default-off invocation refuses analysis before opening inputs; explicit experimental opt-in enables it.
• Match by node, method and operation identity; overlapping starts and malformed records cannot produce a successful timing claim.
• Report input bounds, successful/failed pairs and incomplete evidence, including operations outside the requested window.
• Reject oversize or malformed inputs explicitly, and emit no raw log text or credentials.
• Unit tests reproduce false-completeness cases before the fix, and replay a finite real reader-only sample as an operator.
• Wire tests into required CI, document normal reader collection and the separation from request-level incident attribution.

Success signal and follow-up

Proxy: one reviewed command reproduces the finite timing evidence and reports the known gaps in this round without silently dropping them. Evaluate the real sample before merge. Keep #4329 open until request attribution, repair and recovery are independently established. Reassess the temporary experimental gate after further operator use, by 2026-10-09.

Size: medium, repository-local operator tooling and tests.