Fireworks-hosted models (accounts/fireworks/models/*) display as raw path/slug instead of friendly names

[therickfactr]

Summary

Fireworks-hosted models (id form accounts/fireworks/models/<slug>) display as a
raw path/slug instead of a friendly name, unlike every other provider. In the
codeburn dashboard By Model panel they show as fireworks/mode… (truncated
raw path); in codeburn models they show as glm-5p2, qwen3p7-plus,
kimi-k2p7-code, deepseek-v4-pro, deepseek-v4-flash.

These arrive via the OpenCode provider (an opencode-format session log written
by a Fireworks-backed fleet driver).

Environment

• codeburn 0.9.16 (Homebrew) — note the behavior is the same on current main.
• Provider: OpenCode; models hosted on Fireworks AI.
• macOS.

Root cause

• getShortModelName's path-style fallback returns the last path segment
  verbatim, and there is no SHORT_NAMES entry for these Fireworks slugs, so the
  bare slug leaks (glm-5p2 rather than GLM-5.2).
• Separately, the dashboard By Model panel renders the cached modelBreakdown
  key directly (no getShortModelName pass), so an old/un-normalized cached key
  leaks the full raw path (accounts/fireworks/models/… → truncated fireworks/mode)
  until the session cache is rebuilt.

Proposed fix

Display-only (pricing via getModelCosts is untouched, so no dollar amounts move):

1. getShortModelName path fallback: take the last path segment and re-resolve
   it through the normal name resolution, so a known slug earns a friendly name
   and an unmapped slug still falls through to the raw segment.
2. Add SHORT_NAMES entries for the Fireworks slugs (glm-5p2 → GLM-5.2,
   qwen3p7-plus → Qwen 3.7 Plus, kimi-k2p7-code → Kimi K2.7 Code;
   deepseek-v4-* already resolve).
3. Route the dashboard By Model render through getShortModelName (matching
   overview's "Top models") so already-cached raw keys normalize without a rebuild.

Verified locally against real data: codeburn models now shows GLM-5.2 / DeepSeek
v4 Pro / Kimi K2.7 Code / Qwen 3.7 Plus / DeepSeek v4 Flash with unchanged costs;
tests/models.test.ts extended (the existing last-segment fallback assertion is kept
for a still-unmapped id); full suite shows no new failures.

I have a branch ready and can open a PR if the approach looks right.

[iamtoruk]

Great report, and your root-cause analysis checks out exactly against the code: the path fallback in getShortModelName returns the last segment verbatim by design comment, and the dashboard By Model panel never routes cached keys through getShortModelName at all (overview does, which is why the two disagree). All three parts of your approach are clean, display-only is the right scope, and keeping the last-segment fallthrough for unmapped slugs is exactly the honest behavior we want.

Please open the PR, we'd love to have it. One small heads-up from reading the dashboard path: once cached keys normalize, two raw keys (the full accounts/fireworks/models path from an old cache and the bare slug from a fresh one) can resolve to the same display name, so make sure the By Model aggregation merges those rows into one instead of rendering duplicates. Sounds like your models.test.ts extension already covers the resolution side; a small case for that merge would round it out.

Review will be quick, today has been a shipping day.

[therickfactr]

@iamtoruk, thanks! #787 is ready for review.

I had missed By Model aggregation merges in the first commit, but it is now addressed. I also merged origin/main into the PR branch because it had already moved since my clone this morning -- a shipping day, indeed!