fix(cluster): apply AlphaFloor in dial calibration sweep

[rohith500]

Bug

computeDialCalibration records QualityBias mix-change breakpoints by sweeping unfloored uniform alpha, but live routing applies alpha[i] = max(dialToAlpha(t), floor[i]) via applyDialAlpha.

On floored bundles (v0.72–v0.74, latest = v0.73), roughly the first half of the dial produces an identical model mix / projected cost — the same “50% looks like 20%” failure #479 fixed, reintroduced when #501/#515 added floors without retargeting the calibration sweep.

CI did not catch it: TestRoutingDistribution_NoDeadZone and mid-vs-low cost guards pin on v0.67, which ships no alpha_floor.

Reproduction (v0.73, 21-point RoutingDistribution)

Metric	Before (unfloored calibration)
Breakpoints	30 (9 with α ∈ (0, 0.30))
Cost at dial 0 / 0.2 / 0.5	0.00187 / 0.00187 / 0.00187
Longest identical-mix streak	10 steps

Fix

Extract shared applyAlphaFloor (alpha[i] = max(raw, floor[i]); floor == nil disables) and call it from both:

• computeDialCalibration — sweep raw uniform alpha through the floor before recording mix breakpoints
• applyDialAlpha — request-time / distribution path (same policy as before, now shared)

After the fix on v0.73:

Metric	After
Breakpoints	14 (none below min floor)
Cost at dial 0 / 0.2 / 0.5	0.00187 / 0.00206 / 0.00269
Longest identical-mix streak	1 step

Bundles without alpha_floor (v0.67) keep the previous calibration behavior.

Tests

• TestRoutingDistribution_NoDeadZone_FlooredBundle — v0.73: no long consecutive identical-mix run
• TestRoutingDistribution_MidDialIsPricierThanLowDial_FlooredBundle — v0.73: cost(0.5) ≥ 1.2× cost(0.2) (floored span is tighter than v0.67’s 1.5×)
• TestComputeDialCalibration_NoBreakpointsBelowMinFloor — no interior breakpoint below min(AlphaFloor)
• TestComputeDialCalibration_HeterogeneousFloorSynthetic — different per-cluster floors; no breakpoint in (0, minFloor)
• TestApplyAlphaFloor_NilAndHeterogeneous — nil floor + per-slot floor behavior
• Existing v0.67 NoDeadZone / mid-vs-low guards unchanged (backward-compat path)

Full package: go test ./internal/router/cluster/ -count=1 green.

Closes #779

[greptile-apps]

PR author is not in the allowed authors list.

[devin-ai-integration]

Thank you so much for this contribution, @rohith500 — the fix itself is exactly right, and the floored-bundle regression coverage is genuinely thorough. 🙏

We've rewritten it to match the repo's internal conventions in #788 (your logic and tests are preserved verbatim, with credit to you). The only change: AGENTS.md's comment rule ("Concise comments, sparingly … never reference current task/PR/caller") — comments like // #779: unfloored calibration left a long consecutive dead zone and pre-fix narrative (pre-fix cost(0.5)==cost(0.2)) reference the issue/PR and the old behavior, which we keep out of the code (that context belongs in the PR description instead). In #788 those comments now describe the code's general behavior only.

Please don't take this as a knock on your work — these conventions are internal, evolving, and honestly not obvious from the outside. The engineering here was great: shared applyAlphaFloor helper, non-tautological floored-bundle guards, heterogeneous-floor synthetic — all spot on. Hope to see more PRs from you!