refactor(chat): shared history builder with model/plan/source-aware budgets

[ABB65]

Summary

Studio chat and the Conversation API carried two near-identical copies of the sliding-window history builder: hard-coded 8K token ceiling, magic loadConversationMessages(..., 50) row cap, and divergent tool_calls / toolCalls casing handling. The 8K ceiling was left over from earlier model defaults and is undersized for Claude 4 (200K-window models).

This PR collapses both into server/utils/conversation-history.ts and makes the budget model/plan/source-aware.

Budget table (review-agreed)

// Model — capability + pricing tier
Haiku 4.5            12K
Sonnet 4 / 4.5 / 4.6 32K / 40K / 48K
Opus 4 / 4.1 / 4.7   32K / 32K / 48K
fallback             16K

// Plan — Contentrain's per-message margin posture
free          0      // defensive backstop, gated upstream
starter       0.75x
pro           1x
enterprise    1.25x
community     1x

// Source — who pays for the input tokens
studio / api  1x
byoa          1.5x   // user pays Anthropic directly

maxTokens = base × plan × source. Sonnet/Opus values are conservative on purpose — once prompt caching lands (cache reads cost ~10% of base input) the model table can safely grow. Sources: models overview, pricing.

Changes

• New server/utils/conversation-history.ts — selectHistoryBudget() + buildPromptMessages(). Pure functions; no DB, no provider.
• chat.post.ts — model selection moved before history (model drives budget); IIFE + push loop replaced with two helper calls.
• ee/enterprise/conversation-api.ts — same pattern; duplicate loadConversationMessages wrapper renamed loadConversationHistoryForResponse and now serves only the /history.get route's enriched JSON response shape (the runtime chat path doesn't need usage/createdAt fields).
• Integration test mocks — added vi.mock('~~/server/utils/conversation-history') to chat-route and overage-soft-cap integration files. One ownership-flow assertion relaxed from hard-coded 50 to expect.any(Number) (rowLimit is derived; budget arithmetic covered by the new unit tests).

Test plan

• pnpm test — 605 passed (590 + 15 new unit tests in conversation-history.test.ts)
• pnpm typecheck clean
• pnpm lint — 0 errors on changed files

13 unit tests cover the matrix: per-model budget, fallback, plan multipliers (starter/enterprise/community/unknown), free→0 defensive backstop, BYOA 1.5x, API no-multiplier, rowLimit scaling, empty history, fits-in-budget, exceeds-budget, zero-budget, snake_case tool_calls, camelCase toolCalls, chronological ordering after cutoff.

Out of scope

• Tokenization accuracy. Still using length/4 heuristic — accurate Anthropic countTokens integration is a separate PR.
• Prompt cache breakpoints. Next PR (perf(ai): system-block array + cache_control); once cached, the Sonnet/Opus budget values in the table here can grow with margin headroom.

Net diff

+379 / −64 — most of the size is unit test coverage. Net runtime behavior change: conversations now use 1.5x–6x more history before cutoff depending on model/plan/source, but cost stays bounded by the multiplier table.