fix(agent-sre): calculate latency from windowed measurements - #4207

Description

With a one-hour window, recording 10,000 ms and then 200 ms more than an hour later still reports 10,000 ms. Reconstructing the indicator over an existing SQLite store instead reports no value. Both happen because the percentile reads a separate process-local list instead of the measurement store.

Calculate the percentile from values_in_window() and remove the redundant list. This uses the existing cutoff and persistence behavior without changing percentile indexing or SLO targets.

Validation

Three regression cases fail before the fix: expired outlier, fully expired window, and SQLite restart. Indicator and persistence tests: 65 passed. Full package run: 1,497 passed, 3 skipped, 1 known baseline failure (the separate Windows benchmark clock bug).

All regressions were reproduced on origin/main 406ddd4d before the production change. Whitespace checks pass. Ruff passes on the changed files. Changed-line spelling passes with the existing API identifier Deduplicator allowed in the temporary check input.

Scope: the existing Agent-SRE source tree, which remains exercised by repository CI. Combined validation of these six independent SRE fixes: 1,505 passed, 3 skipped on Windows/Python 3.12.

Type of Change

• Bug fix (non-breaking correction of existing behavior)

Package(s) Affected

• agent-sre

Checklist

• Changed code follows the project style checks; the exact validation scope is recorded above.
• Regression tests or existing package tests verify the change, as detailed above.
• The independent full package run has no failures: one pre-existing Windows benchmark failure remains, as documented above. Focused tests pass; the combined run with fix(agent-sre): measure benchmark duration with a monotonic clock #4206 passes (1,505 passed, 3 skipped).
• Documentation is updated where needed; the behavior and compatibility implications are described above.
• The Microsoft CLA check has passed for this contribution.

Attribution & Prior Art

• Related prior contributions are credited below; this PR introduces no unattributed code from external projects.
• External design influences, where applicable, are identified below.

Prior art / related projects: Uses the existing Agent-SRE implementation and standard library operations. No code was copied or adapted from an external project.

AI Assistance

I reviewed the specific changes in this diff before this PR was opened.

Codex assisted with implementation, regression tests, and local validation under my direction.

• I reviewed this diff before submission; its behavior, rationale, and tradeoffs are documented above.
• Tests and verification appropriate to this change were run, with limitations disclosed above.
• Submission followed my direction and my review of these specific changes.
• No AI-generated reviews of other contributors' PRs were posted as part of this contribution.

IP, Patents, and Licensing

• This change uses the existing project implementation and ordinary software techniques; no proprietary or patent-specific technique is introduced.
• Understanding or using this change requires no NDA or separate licensing agreement.
• This contribution remains under the repository's MIT License and introduces no third-party source code or additional license terms.

Related Issues

Newly reproduced defect; the reproduction and regression coverage are described above.