fix(cluster): computeDialCalibration ignores AlphaFloor — QualityBias dial dead zone on floored bundles

[rohith500]

Summary

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 current 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 class of “50% looks like 20%” failure that #479 fixed, reintroduced when #501/#515 added floors without retargeting the calibration sweep.

Why it matters

• Routing correctness / cost-quality UX on the customer-facing QualityBias dial (dashboard + x-weave-routing-*).
• Demonstrated on committed production artifacts, not a synthetic edge case.
• Availability unaffected; quality and spend mix stop responding to dial travel until ~t=0.5.

Where

• internal/router/cluster/scorer.go — computeDialCalibration (unfloored sweep)
• internal/router/cluster/scorer.go — applyDialAlpha / dialToAlpha (floored consumer used by Route and RoutingDistribution)

Reproduction (v0.73)

Same 21-point RoutingDistribution grid used by #479’s regression tests:

Bundle	alpha_floor	Breakpoints with α ∈ (0, minFloor)	Adjacent identical mixes	Cost at dial 0 / 0.2 / 0.5
v0.67	none	0	1	0.00010 / 0.00033 / 0.00116 ✅
v0.70–0.71	0.30	1	4	mild degradation
v0.72–0.74 (latest)	0.30	9	11 (max run 10)	0.00187 / 0.00187 / 0.00187 ❌

Counterfactual: recompute calibration with per-cluster max(a, floor[i]) (same policy as applyDialAlpha) → breakpoints 30 → 14, zero sub-floor breakpoints.

Historical regression mechanism

1. fix(routing): calibrate quality dial against real mix crossovers #479 — calibrated dial against real mix crossovers (computeDialCalibration / dialToAlpha). Regression tests (TestRoutingDistribution_NoDeadZone, mid-vs-low cost) pin on v0.67, which ships no alpha_floor.
2. fix(routing): per-cluster alpha_floor so the price/quality dial keeps every cluster at its best-for-budget model #501 — added alpha_floor via applyDialAlpha = max(dialToAlpha(t), floor[i]). Did not update the calibration sweep. Reviews did not mention computeDialCalibration.
3. fix(routing): gate agentic pool to harness-capable models so the price dial can demote Opus #515 — relaxed agentic floors 0.88 → 0.30 and moved harness guard to the candidate pool. Calibration still unfloored.
4. v0.72+ — roster/fable changes multiplied unreachable low-α breakpoints (1 → 9); dead zone became customer-visible on latest.

Recommended fix

Share a helper that turns a raw uniform alpha into the effective floored per-cluster vector, and use it from both computeDialCalibration and applyDialAlpha.

Regression coverage:

• Synthetic heterogeneous-floor test (different floors per cluster).
• Frozen floored-bundle regression (v0.73 or successor): no long identical-mix run; mid dial materially pricier than low.
• Backward-compat: bundles without alpha_floor keep current calibration behavior.
• Before/after RoutingDistribution / cmd/routing-report --quality-bias report.

Do not keep NoDeadZone / mid-vs-low guards only on v0.67 — that is how this regressed silently.

Confidence

High — reproduced on committed artifacts; counterfactual confirms root cause; no open issue tracks this.

[rohith500]

Post-filing verification (re-ran thoroughly)

Re-checked before treating this issue as locked. All claims hold.

Live Route path (not only distribution preview)

On v0.73 with a fixed embedding, QualityBias → applyDialAlpha → same floored alphas as RoutingDistribution:

dial t	dialToAlpha	α[0] (floor 0.30)	α[2] (floor 0.55)	Route model
0.00	0.0000	0.300	0.550	gemini-3.5-flash
0.25	0.2469	0.300	0.550	gemini-3.5-flash
0.50	0.3988	0.399	0.550	gemini-3.5-flash
0.75	0.5513	0.551	0.551	gemini-3.1-pro-preview
1.00	1.0000	1.000	1.000	claude-fable-5

has_tools: true does not remove the dead zone (same models at t=0/0.25/0.5); at mid dial conversational clusters remain floor-pinned (dialToAlpha(0.5) ≈ 0.40 < 0.55).

Distribution metric (cross-cluster)

• cost[0] = cost[0.5] = 0.00187; cost[1] = 0.00606
• Unfloored breakpoints: 30, of which 9 have α ∈ (0, 0.30)
• Floored counterfactual breakpoints: 14

CI blind spot confirmed

Bundle	identical adjacent mixes (21-grid)	Would fail NoDeadZone (≤3)
v0.67 (no floor; what CI tests)	1	no
v0.73 (latest, floored)	11	yes

Root cause unchanged: calibration sweep sets alpha[i] = a with no floor; live path uses max(dialToAlpha(t), floor[i]).