fix(proxy): wire no-progress detector into OpenAI chat completions

[rohith500]

Summary

Fixes #825: ProxyOpenAIChatCompletion now runs the same cross-envelope no-progress detector that ProxyMessages already has (5 identical dispatch fingerprints within 90s → synthetic stop via handleNoProgressBreak).

• Reuses existing computeNoProgressFingerprint / recordAndDetect / handleNoProgressBreak — including the previously production-dead FormatOpenAI synthetic chat.completion branch.
• Gates match Messages: !agentShadowMode && !AuthoritativePerTurn && toolBearingTurn && noProgress != nil.
• OpenAI tool-bearing detection uses the same envelope helpers (AssistantToolCallSignatures + LastUserMessage().HasToolResult for trailing role=tool messages), snapshotted before compaction like Messages.
• ProxyOpenAIResponses still delegates to ProxyOpenAIChatCompletion (confirmed), so Responses ingress inherits this automatically. Gemini out of scope.

Test plan

• Commit 1 (red): TestProxyOpenAIChatCompletion_NoProgressBreaksAtThreshold failed with brokeAt=0, upstreamHits=6
• Commit 2 (green): same test passes; breaks on attempt 5 with No-progress loop detected
• TestHandleNoProgressBreak_OpenAIChatCompletionShape — asserts object=chat.completion, assistant/stop, prompt_tokens
• Negative control: varying tool args ×6 — no break, 6 upstream hits
• Existing TestNoProgress* suite still green (Messages path unchanged)
• make check passed
• Adversarially verified (not committed): text-only identical ×6 no
  break, 4-below-threshold no break, 5-across-expired-window no break,
  same-messageCount-different-content no false collision, 16-concurrent
  no panic

Before / after (positive control)

Before (commit 1 only):

--- FAIL: TestProxyOpenAIChatCompletion_NoProgressBreaksAtThreshold
    Error: Not equal: expected: 5 actual: 0
    Messages: ProxyOpenAIChatCompletion must break on the 5-th identical fingerprint (got brokeAt=0, upstreamHits=6)

After (commit 2):

=== RUN   TestProxyOpenAIChatCompletion_NoProgressBreaksAtThreshold
... "message":"No-progress loop detected; breaking turn" ... "repeat_count":5 ...
--- PASS: TestProxyOpenAIChatCompletion_NoProgressBreaksAtThreshold

Follow-up (non-blocking)

With no pin store configured, both Messages and OpenAI fall back to a
shared install:<id>:<role> fingerprint bucket (pre-existing coarseness
for Messages alone; now also shared with OpenAI). In production with a
pin store, keys are per-session via DeriveSessionKey, so this only
matters for installations running without one. Not a regression this PR
introduces — the install-bucket fallback already existed — just newly
shared across two surfaces. Worth a follow-up look if pin-store-less
installations turn out to be common enough to matter; not blocking this
fix.

[greptile-apps]

PR author is not in the allowed authors list.

[devin-ai-integration]

Thanks for this, @rohith500 — really clean contribution. Reviewed against the root AGENTS.md/CLAUDE.md and internal/proxy/CLAUDE.md, and it follows the repo conventions nicely, so no changes needed on my end:

• Right layer. The detector is wired on *proxy.Service.ProxyOpenAIChatCompletion (the routing/dispatch home per internal/proxy/CLAUDE.md), with no new I/O — it reuses the existing computeNoProgressFingerprint / recordAndDetect / handleNoProgressBreak machinery rather than duplicating it (DRY, root AGENTS.md).
• Faithful parity with ProxyMessages. Same gate (!agentShadowMode && !AuthoritativePerTurn && toolBearingTurn && noProgress != nil), same tool-bearing detection via AssistantToolCallSignatures + trailing role=tool HasToolResult, snapshotted before compaction — a good instance of "put logic in the package that uses it" without re-inventing it.
• No magic strings in prod code — routing goes through catalog.TierFor/roleForTier, and the proxy_test file uses providers.ProviderOpenAI/providers.ProviderAnthropic constants (root AGENTS.md → Conventions → Go style).
• Non-tautological tests — the positive control asserts the exact break attempt + upstream hit count + chat.completion JSON shape, and the negative control (varying tool args) proves it doesn't trip; both would fail if the prod wiring were removed (root AGENTS.md → Tests). Nice touch keeping them as in-memory fake-repo tests rather than DB-backed integration tests in internal/.
• Comments are concise/why-focused and the #825 guard refs match the established convention in this package (cf. conformance_responses_test.go #331, enabled_providers_internal_test.go PR #159).

CI is green. LGTM — thanks again!