fix(model): keep callback data short

[fengjikui]

Summary

• keep /models favorite/recent callbacks short by using list references instead of embedding provider/model IDs
• keep model search result callbacks short by storing the selected models in interaction metadata
• add regression coverage for long Fireworks-style model IDs that exceed Telegram callback_data limits

Why

Telegram rejects inline keyboard callback_data values longer than 64 bytes. Long provider/model IDs such as Fireworks deployment paths could make the model menu fail with BUTTON_DATA_INVALID before the user could select anything.

Fixes #178.

Validation

• npx prettier --check src/bot/menus/model-selection-menu.ts src/bot/callbacks/model-selection-callback-handler.ts tests/bot/handlers/model.test.ts
• npm test -- --run tests/bot/handlers/model.test.ts
• npm run build
• npm run lint
• npm test
• git diff --check

[grinev]

@fengjikui thanks for the fix. One issue before merging:

Stale clicks on result buttons can select a wrong model. After the model-search interaction is cleared (e.g. by the next interactionManager.start()), clicking an already-rendered model:result:0 button reaches handleModelSelect and parseLegacyModelCallback parses it as providerID="result", modelID="0" — and the bogus model gets selected. Pre-PR, the real provider/model were embedded in the callback_data, so stale clicks still worked.

Same shape for model:list:recent:0 reaching handleModelSearchResults (the messageId guard mitigates it, but the legacy branch would otherwise read it as providerID="list", modelID="recent:0").

Could you make the legacy fallbacks reject the new prefixes? E.g. in handleModelSelect, explicitly drop/log model:result: instead of falling through to parseLegacyModelCallback, and exclude model:list: from the fallback in handleModelSearchResults. That way stale clicks fail safe instead of selecting a wrong model.

[fengjikui]

Thanks for catching that. I pushed c6f1b53 to make the new short callback prefixes fail safe instead of falling through to the legacy parser:

• model:result:* is now rejected by handleModelSelect when the search interaction is no longer active, so stale result buttons no longer become providerID="result" / modelID="0".
• unresolved model:list:* callbacks no longer fall back to legacy parsing in handleModelSelect.
• handleModelSearchResults also rejects short model:list:* / model:result:* prefixes from its legacy fallback path.

Added regression coverage for stale model:result:0, unresolved model:list:recent:0, and the search-results fallback case.

Validation:

• npm test -- tests/bot/handlers/model.test.ts (19 tests)
• npx prettier --check src/bot/callbacks/model-selection-callback-handler.ts tests/bot/handlers/model.test.ts
• npm run build
• npm run lint
• git diff --check
• npm test (121 files / 1069 tests)

[fengjikui]

Self-review found one more index-stability edge case: model:list callbacks were re-reading model.json at click time, so a favorites/recent reorder after the menu was rendered could retarget an existing button. f05fd04 now stores the rendered lists in the active inline-menu metadata and resolves short callback indices from that snapshot. Validation: targeted model/inline-menu tests (28 tests), full suite (122 files / 1096 tests), build, lint, Prettier, and git diff --check.

[grinev]

@fengjikui thanks for contribution!