fix(routing): per-cluster alpha_floor so the price/quality dial keeps every cluster at its best-for-budget model

[steventohme]

Root cause

The QualityBias dial resolved to a single uniform alpha, overwriting the bundle's shaped per-cluster alpha vector (v0.70: 0.96 on coding/agentic/knowledge, 0.8 on conversational). That shape exists so a price-leaning dial never cheaps out agentic turns — but flattening it discarded the protection. A mid dial maps to ≈0.60 effective alpha; at 0.60 the cheapest "good-enough" model wins every cluster. In prod this routed a /cloud-dbg skill turn to minimax-m3, which can't drive the Claude Code tool/skill protocol (it grepped the filesystem for the skill instead of running it).

Fix: declarative per-cluster alpha_floor

The dial now floors each cluster at a bundle-declared minimum:

alpha[i] = max(dialToAlpha(t), floor[i])

alpha_floor is an optional per-cluster vector shipped in default_routing_knobs — the lowest quality weight the bundle tolerates per cluster at maximum price-sensitivity. It is data, not code: no magic constant, no guessing which clusters to protect. nil/absent ⇒ legacy uniform dial (no behavior change for bundles that don't ship it, e.g. v0.67). Validated at load (length == K, each in [0,1]). Shared by Route and RoutingDistribution so the dashboard preview matches live routing.

Supersedes the first revision of this PR, which hardcoded a 0.92 constant + inferred protected clusters from the alpha shape. That coupled routing policy to scorer code and was fit to one corpus.

v0.70 floor vector

• Coding / agentic / knowledge clusters → 0.88 (stay on premium models at any dial)
• Conversational / trivial clusters → 0.55 (still far cheaper than their 0.80 default, but not collapsed to the cheapest model — best quality for their budget)

Verification

New --quality-bias flag on cmd/routing-report. On the v0.70 corpus at quality_bias=0.2:

register	before (uniform dial)	after (alpha_floor)
agentic_tool	0% premium → cheapest (minimax/flash)	53% opus + 35% glm-5.2 (no minimax)
hard_code	0% premium	40% opus / 57% glm-5.2
knowledge_qa	low	53% opus
conversational	cheapest model	deepseek-v4-pro / mimo (best-for-budget)

Protected clusters are stable across the low half of the dial (floor pins them), then climb to ~all-opus at the quality end. Conversational stays cheap but never collapses to the bottom model.

Tests

• TestApplyDialAlpha_HoldsEachClusterAtItsDeclaredFloor — every cluster pinned at its declared floor at t=0; reaches 1.0 at t=1.
• TestApplyDialAlpha_NilFloorIsUniformDial — no floor ⇒ legacy uniform dial.
• TestApplyDialAlpha_AgenticStaysOffCheapModelAtLowDial — agentic centroid moves off the cheap model to a frontier model at a low dial.
• Existing v0.67 distribution/dial tests unchanged (ships no alpha_floor).

Notes / open questions

• In-place edit of v0.70 metadata. alpha_floor is added to v0.70's default_routing_knobs (hand-maintained routing config, like the existing shaped alpha). It's additive and changes no clustering/quality data, but it does alter a same-day-shipped deployed version in place — flag if you'd prefer a v0.71 bump.
• Known residual: prod runs top_p=2 (a cluster blends with its nearest neighbor), so an agentic cluster paired with an unprotected neighbor can partially cheapen at the extreme low dial. The reported {0,4} case (both protected) is fully fixed.
• Floor values are now trivially tunable in metadata — raise the 0.88 entries if you want agentic even more opus-heavy at a price dial.

[greptile-apps]

_{Reviews (1): Last reviewed commit: "fix(routing): floor protected clusters s..." | Re-trigger Greptile}

[github-actions]

🧭 Cluster routing review

ℹ️ Informational — candidate artifact (latest unchanged); no sign-off required.

Cluster routing report — v0.70

Routed 180 labeled probes through the real Scorer at top_p=2, embedder jina-v2-base-code-int8. Premium = opus-4-7/4-8, gpt-5.5, gemini-3.1-pro, sonnet-4-6.

Register → model (this artifact)

register	Opus %	premium %	top models
conversational	33%	43%	deepseek-v4-pro:16 claude-opus-4-8:10 gemini-3.1-pro-preview:2 gemini-3.5-flash:1
trivial_nl	13%	50%	deepseek-v4-pro:13 gemini-3.1-pro-preview:10 claude-opus-4-8:4 gpt-5.5:1
knowledge_qa	70%	77%	claude-opus-4-8:21 glm-5.2:4 deepseek-v4-pro:2 gemini-3.1-pro-preview:1
easy_code	73%	87%	claude-opus-4-8:22 deepseek-v4-pro:3 gemini-3.1-pro-preview:2 gpt-5.5:2
hard_code	60%	80%	claude-opus-4-8:18 gpt-5.5:6 glm-5.2:6
agentic_tool	87%	93%	claude-opus-4-8:26 gpt-5.5:2 glm-5.2:2

What each cluster represents (probe register mix by nearest centroid)

cluster	probes	dominant	register breakdown
0	36	agentic_tool	agentic_tool:12 knowledge_qa:9 hard_code:8 conversational:4 easy_code:2 trivial_nl:1
1	8	easy_code	easy_code:3 agentic_tool:2 conversational:1 knowledge_qa:1 trivial_nl:1
2	2	conversational	conversational:2
3	1	knowledge_qa	knowledge_qa:1
4	3	conversational	conversational:1 easy_code:1 trivial_nl:1
5	14	agentic_tool	agentic_tool:5 conversational:2 easy_code:2 hard_code:2 knowledge_qa:2 trivial_nl:1
6	1	easy_code	easy_code:1
7	6	agentic_tool	agentic_tool:2 easy_code:2 hard_code:1 knowledge_qa:1
8	9	knowledge_qa	knowledge_qa:4 agentic_tool:2 hard_code:2 easy_code:1
9	10	conversational	conversational:4 trivial_nl:4 easy_code:1 knowledge_qa:1
10	24	conversational	conversational:12 trivial_nl:7 easy_code:3 knowledge_qa:2
11	4	trivial_nl	trivial_nl:2 easy_code:1 knowledge_qa:1
12	13	easy_code	easy_code:9 knowledge_qa:2 agentic_tool:1 trivial_nl:1
13	25	hard_code	hard_code:17 agentic_tool:6 easy_code:1 knowledge_qa:1
14	7	knowledge_qa	knowledge_qa:4 easy_code:2 trivial_nl:1
15	17	trivial_nl	trivial_nl:11 conversational:4 easy_code:1 knowledge_qa:1