feat(sync): add Fireworks AI catalog sync

[opencode-agent]

Summary

• add an authenticated fireworks-ai sync target using Fireworks' OpenAI-compatible /inference/v1/models endpoint
• update existing entries only from safe API signals, retain account-scoped absences, and report new generation IDs for manual completion
• preserve authored pricing, output limits, reasoning controls, and exact context caps below Fireworks' advertised ceiling
• wire the provider into hourly sync CI with its dedicated repository secret
• add Fireworks image input to qwen3p8-max, confirmed by both the API and a successful image completion

API investigation

• The dedicated account currently sees 25 callable entries: 22 model resources and 3 routers. Two model resources are embedding/reranking entries, leaving 23 generation entries.
• The repository already tracks 21 generation entries. The two untracked generation IDs are accounts/fireworks/models/minimax-m2p7 and accounts/fireworks/models/deepseek-v4-pro.
• format=openrouter does not produce OpenRouter metadata. It returns a byte-identical response to the default endpoint, as do type=text, limit=1, and pageSize=1.
• /v1/accounts/fireworks/models reports 306 public model resources, but filter=supports_serverless=true returns only 21 base models. It excludes all 3 callable routers and the callable minimax-m2p7 model.
• The management response exposes richer descriptive fields but returned empty serverlessModes, so it supplied no advertised pricing or serving-path data.
• The separate routers endpoint returned 403 for an ordinary inference credential.
• Capability booleans can have false negatives: minimax-m3 reports supports_image_input=false but successfully accepted an image completion. The sync therefore uses positive image/tool flags to add capabilities but never removes authored capabilities from a false flag.

Checks

• bun test packages/core/test/fireworks-ai-sync.test.ts — 5 passed
• bun validate — passed
• live bun models:sync fireworks-ai --dry-run --no-issues — clean after the initial update (21 unchanged)
• full bun test retains the same 5 unrelated failures reproduced on clean origin/dev

Requested by: @rekram1-node (Aiden via Slack)

[github-actions]

No actionable findings.

[deepsimulation]

Could we use GET /v1/serverless/models as the sync source instead of /inference/v1/models? It exposes serverless-specific pricing, serving modes, aliases/usage identifiers, context, and modalities.

I'm happy to open up new PR.

[rekram1-node]

closing in favor:
#7112