Pricing matcher has no general normalizer for vendor-renamed auto-pilot/canary model ids (#187 was patched per-name)

[pitimon]

Follow-up to #187/#188.

Problem

#188 fixed claude-auto-pilot-fable-v1-canary / claude-mythos-*-canary by adding two literal curated.fuzzy substring rules ("fable", "mythos"). This is a per-model-name patch, not a fix to the underlying gap: src/lib/pricing/matcher.js has no tier that recognizes the <vendor>-auto-pilot-<name>-v<N>-canary naming shape in general.

Any future renamed/canary model whose base name is not already a short LiteLLM substring (the reason gpt-5.6-auto-pilot-* survived — "gpt-5" / "gpt-5.6" happen to already be short LiteLLM keys) will reproduce the exact same silent-$0 bug, requiring another one-line curated patch and another release cycle.

Also a real risk introduced by the current fix: {"match": "fable", "ref": "claude-fable-5"} is a bare substring rule — any future model id containing the literal substring "fable" (e.g. a hypothetical distinct cheaper fable-mini) will silently inherit the Opus-tier ($10/$50) Fable 5 price via fuzzy match, not because it is actually that model, but because the substring matched.

Ask

Design a normalizer for source: "hermes" (or general) that strips known infix/suffix noise (-auto-pilot-, -v\d+-canary, -canary, -preview) before running the existing lookup ladder, so vendor-renamed ids resolve to their real base model without a hand-added fuzzy rule per name. This is matcher.js resolve-order work — evaluate as a proper design pass, not another one-line patch, given the risk noted above.

Evidence

• src/lib/pricing/matcher.js:103-206 (lookupPricing resolve order, all 7 tiers require contiguous substring or exact match)
• src/lib/pricing/curated-overrides.json (fable/mythos fuzzy rules added in fix(pricing): resolve fable/mythos canary variants via fuzzy match #188)
• Repro: claude-auto-pilot-fable-v1-canary and claude-mythos-v1-canary both hit miss pre-fix(pricing): resolve fable/mythos canary variants via fuzzy match #188; gpt-5.6-auto-pilot-*-canary survived only because litellm:fuzzy happened to match the short gpt-5/gpt-5.6 LiteLLM keys as substrings.

[pitimon]

Resolved in #194, released as @ipv9/tokentracker-cli@0.39.54.

• matcher.js gained stripVendorRenameNoise() + a scoped denoise retry: strips -auto-pilot- infix and trailing -vN-canary/-canary, retried through the full lookup ladder exactly once, only after every other tier misses on the raw id. A future auto-pilot/canary rename of an already-priced model (e.g. claude-auto-pilot-opus-4-8-v1-canary) now resolves via litellm:exact/curated:exact with no curated fuzzy rule needed.
• Also fixed the bare-substring collision risk this issue flagged: narrowed #188's fable/mythos fuzzy rules to claude-fable/claude-mythos. Verified against the live pricing seed that claude-mythos-preview (a real, distinct LiteLLM-priced model) was one design change away from being silently mispriced by the bare rule — confirmed it still resolves correctly (litellm:exact) after narrowing.

Verified: node --test test/pricing.test.js (39/39, 5 new regression tests), full npm run ci:local gate, re-checked against the fresh 0.39.54 pricing seed post-publish.