Recalibrate warm P95 budgets after schema-v2 baselines

[karthiknadig]

Problem

PR #510 changes the existing full_refresh and time_to_first_env distributions from one cold + nine warm samples to ten independently paired warm samples. The current P95 budgets were calibrated against the old mixed distribution and will become too permissive once schema-v2 baselines land.

Three schema-v2 hosted runs report:

Platform	Warm refresh P95	Warm time-to-first P95	Current absolute budgets
Linux	60 / 58 / 69ms	15 / 16 / 15ms	1,000ms / 250ms
Windows	171 / 143 / 206ms	38 / 28 / 47ms	5,000ms / 500ms
macOS	198 / 270 / 171ms	166 / 228 / 136ms	1,000ms / 750ms

After #510, those old limits could allow multi-second warm-cache drift even though cold performance is gated separately.

Scope

After #510 merges and exact schema-v2 main baselines exist:

• inspect repeated warm refresh/time-to-first baseline variance on all platforms;
• tighten warm P95 dual budgets with measured hosted-runner headroom;
• keep P50, server-startup, and cold-refresh budgets unchanged unless independent evidence requires adjustment;
• update comparator tests and docs/QUALITY_SNAPSHOTS.md;
• retain the schema-v1 migration behavior only where still required.

Acceptance criteria

• Repeated unchanged-head schema-v2 runs stay green.
• Synthetic multi-second warm refresh and warm time-to-first regressions fail on every platform.
• Cold P50 gating remains independent and unchanged.
• Missing metrics, inventory mismatches, and malformed snapshots still fail closed.
• Full CI, coverage, performance, and CodeQL workflows pass.

Depends on #510.

[karthiknadig]

Exact merged schema-v2 baseline run 31516981874 passed at ad7ca14. Combined with the three unchanged-code PR runs:

Platform	Warm refresh P95 range	Warm time-to-first P95 range
Linux	58-69ms	15-16ms
Windows	109-206ms	24-47ms
macOS	160-270ms	133-228ms

Proposed dual budgets:

Platform	Warm refresh P95	Warm time-to-first P95
Linux	50ms / 50%	25ms / 100%
Windows	250ms / 100%	100ms / 100%
macOS	300ms / 100%	250ms / 100%

Against the merged baseline these block above roughly 110ms/41ms on Linux, 359ms/124ms on Windows, and 460ms/383ms on macOS. Absolute headroom is at least 2.5x each observed four-run range while reducing old effective ceilings by 48-95%.

Cold P50, P50, and server-startup budgets remain unchanged. Tests will prove each observed worst run passes and multi-second warm regressions fail on every platform.

[karthiknadig]

Final quality conclusion for PR #512 at 135cb20:

• Exact-head Copilot review: 3/3 files, 0 comments, 0 unresolved threads.
• Full CI, lint, CodeQL, Linux/Windows coverage, and all performance jobs pass.
• Linux warm refresh/time-to-first P95: 53/12ms vs 60/16ms baseline.
• Windows rerun warm P95: 122/23ms vs 109/24ms baseline.
• macOS warm P95: 150/111ms vs 160/133ms baseline.
• All artifacts contain 10 samples, matching environment/manager inventories, schema v2, and empty interpreter-timeout maps.
• Line/function coverage is unchanged on Linux and Windows.
• Comparator suite: 33/33 passed; exact schema-v2 baselines pass and synthetic multi-second warm regressions fail on every platform.
• The first Windows attempt failed only the unchanged warm P50 gate (164ms vs 105ms baseline); the isolated rerun passed at 117ms. That calibration issue is tracked separately in Recalibrate Windows warm P50 budget for schema-v2 variance #513 rather than widening test: tighten schema-v2 warm P95 budgets (Fixes #511) #512's P95 scope.