fix: trim conversation history to fit context window (prevents 'No response came')

[Fahad090NP]

Summary

Fixes two related "No response came" failure modes with a single root-cause fix:

1. Multiple VS Code windows — chatting in separate windows intermittently fails with no response.
2. Repeated failures fixed by Compact Conversation — after many turns the request fails every time; running Compact Conversation and sending one message works again.

Root cause

provideLanguageModelChatResponse sent the entire conversation history as the request payload. The only history guard was image trimming (trimOldImagesFromHistoryInPlace) — there was no text-history truncation. So a long conversation grows past the model's context window and the upstream either:

• rejects the oversized request with HTTP 400 (empty assistant message, finish_reason: null), or
• returns an empty stream → VS Code surfaces "No response came".

Compact Conversation shrinks the history, which is exactly why it "fixes" the symptom. (This also explains the reported Muse Spark ~990 KB payload — it is the uncompacted history, not a model-specific quirk.)

Changes

• src/provider/historyTrim.ts (new, pure) — trimOldMessagesToFitContext drops the oldest messages until the estimated prompt tokens fit contextWindow − maxOutputTokens − safetyMargin. It preserves the first (system/anchor) and last (current prompt) messages, and never splits a tool-call group (stops before the first assistant message carrying tool_calls), so no tool reference is orphaned.
• src/provider/OpenCodeProvider.ts — calls it right after the image-history trim, before the prompt-token estimate, with a [history-trim] diagnostic log.
• src/config.ts — HISTORY_TRIM_SAFETY_MARGIN_TOKENS = 2048.
• Tests — src/test/messages.test.ts (no-op when fitting, drops oldest keeping anchor + last, never splits a tool group, no-op when only anchor + last remain).
• Docs — CHANGELOG Fixed entry + docs/issues/68-20260820-history-trim-context-overflow.md.

Verification

• npm test — 337/337 pass
• npm run lint — all 7 gates pass (Editorconfig, ESLint, Markdown, Prettier, Shell, TypeScript, Tests)

PEACE BE UPON YOU

[Fahad090NP]

Superseded by #169 — that PR is branched from this one and adds the hardening (70% context ratio, 512 KB byte cap, O(n) trim, tool-group dropping) that resolves the 503 / 10-min timeout / empty-response / session-slowness symptoms and closes #165.