fix(validation): honor segment sample percentage on large releases (#812)

[javi11]

Fixes #812.

Problem

The segment scan percentage setting had no effect below 100%. Sampling applied a hard 55-sample ceiling after the percentage math, so on any release with more than ~1,100 segments every setting from 1% to 99% collapsed onto the same 55 STATs. Only 100% behaved as configured, because it takes an early-return path that skips sampling entirely — which is exactly the asymmetry the reporter observed.

The UI presents this as a plain percentage slider ("1% (FAST)" → "100% (SLOW)") with no mention of a cap, so the ceiling was silently overriding an explicit user choice.

Changes

• internal/usenet/validation.go — selectSegmentsForValidation (health checks; the path the settings slider drives) drops min(..., 55) and honors the configured percentage exactly. The min-5 floor for statistical validity is kept.
• internal/importer/validation/fast_fail.go — selectFastFailSegments (import fast-fail, per file) had the same maxSamples = 55 ceiling, which capped even 100% sampling. It now takes first + last + exactly pct% of the middle; the redundant clamp is removed and slice capacity is sized from the real sample count.
• Stale ≤55 comments in processor.go and the FastFailReleaseProbe doc updated.

Reviewer notes

This raises default sampling cost. The 55 cap was doing real work to keep out-of-box checks cheap. At the 5% health default, a ~7,000-segment release moves from 55 STATs to ~350 per check; the 1% import default is milder. That is the intended meaning of the setting, but if it proves too aggressive in practice, lowering the shipped default in manager.go:1663 is a one-line follow-up. The alternative considered was making the cap a new config knob — rejected as extra surface for a setting that should already mean what it says.

Testing

The two tests that asserted the 55 cap are rewritten as #812 regressions, and were confirmed red against the old code (got 55, want 2000) before the fix:

• health sampling: 20,000 segments → 2,000 at 10%, 10,000 at 50%, strictly monotonic across percentages.
• fast-fail sampling: 10,000 at 100%, 1,002 at 10%, no duplicates.

go build ./..., go vet ./internal/... and the full go test ./... pass; internal/usenet, internal/importer, internal/importer/validation and internal/health re-verified with -race.