fix(models): resolve and merge raw ids in models report

[avs-io]

Problem

codeburn models shows raw ids (gpt-5.6-sol, accounts/fireworks/models/kimi-k2p6) that codeburn report already resolves, and splits one model into two rows when two raw ids are the same SKU.

Closes #967.

Root cause

models asked each provider for a label, then bucketed by raw id. Providers with a local table miss fell back to the raw id. A later merge keyed on the display string, so live label collisions (GPT-5 / GPT-5.3 Codex / Kimi K2 Thinking / Opus 4.6) would sum distinct SKUs into one row.

Change

• Provider-first resolution stays. Local labels (Cursor estimated suffix, provider overrides) still win.
• If a provider echoes the raw id, fall back to the global short-name table.
• cursor-agent and vercel-gateway consult that table after their local transform (suffix / vendor peel).
• Merge by provider + alias-resolved canonical id. Display names stay cosmetic.
• Keep the first-seen raw id (models --format json is not a lex-smallest rewrite).
• Mixed rated/unrated Codex credit merges partial-sum the rated buckets and set creditsIncomplete.

User impact

codeburn models matches report for gpt-5.6-* and Fireworks path-form ids. Path-form glm-5p2 merges with the bare slug. Distinct SKUs that only share a label stay two rows.

Preservation

No new LiteLLM rates. No desktop / By Model rewrite (aggregateModelTotals already merges). No Keychain / Buzz / #868 / #916 / #998 work. Audit still buckets by raw id (it recomputes rates per id). No second Extra High.

Testing

• 40 models-report tests + resolveCanonicalModelId unit cases (230 passing in the focused files; 3 pre-existing DeepSeek snapshot fails)
• Maintainer cases: provider-table miss (gpt-5.6-sol → GPT-5.6 Sol (est.)), path-form (accounts/fireworks/models/kimi-k2p6 → Kimi K2.6), canonical merge of fireworks glm-5p2 + glm-5p2, display-only collisions stay split, first-seen raw id, mixed credit partial-sum
• Same display name across providers stays two rows

[iamtoruk]

Reviewed in depth on a rebased branch with real-data runs. The display-resolution half is right and verified ($-identical on a 31-row real corpus, gpt-5.6-sol → "GPT-5.6 Sol" etc.). Requesting changes on the merge half:

Blocker — the merge key is a display string. bucketKey(provider, modelDisplayName, …) merges rows whose labels collide, and there are live same-provider collisions today: "GPT-5" ← gpt-5 | gpt-5-fast | gpt-5.2-low; "GPT-5.3 Codex" ← gpt-5.3-codex | gpt-5.1-codex-high | warp-auto-efficient; "Kimi K2 Thinking" ← kimi-k2-thinking | kimi-code | kimi-for-coding; "Opus 4.6" ← claude-opus-4.6 | warp-auto-powerful. Those are separate SKUs (subscription vs API, fast tier vs base) whose spend would sum into one row. grok-4.5-build stays split only because #1029 hand-renamed its label — this converts every future display-name collision from a cosmetic bug into a money bug. Key the merge on the alias-resolved canonical id; keep display cosmetic.

Also:

• High: if (bucket.model < existing.model) existing.model = bucket.model makes models --format json emit the lexicographically-smallest raw id after a merge — an undocumented payload-semantics change (README promises exact ids for --unpriced).
• Medium: merging a credits-rated bucket with an unrated one nulls credits, which usage-aggregator sums into the menubar codexCredits — a merge can zero a real credit total. Partial sum + incomplete flag instead.
• Cleared: the provider-file changes are display-only (no cache/dedup key impact — correctly no parse-version bump). Pre-existing note: BUILTIN_ALIASES resolves before SHORT_NAMES, so the 'gpt-5.2-low' display entry is dead — that mechanism is what makes the collisions above live.

Suggest splitting: land fallbackRawModelDisplayName now; re-key the merge + drop the row.model mutation in round two. Sequencing note: this should land after #1050 (its display-name regexes match getShortModelName output this PR changes).

[ozymandiashh]

The original canonical-ID merge work looks sound, but this head cannot be merged safely against current main. Its conflicted report paths still compute billable output as output + reasoning; current main deliberately uses billableOutputTokens because Codex output already includes reasoning. Taking the branch side would reintroduce double counting (for example 500k output + 500k reasoning becomes 1M instead of 500k). Please rebase, retain the current billableOutputTokens pricing semantics in both audit and models reports, and rerun the report/model suites on the resolved tree.