fix(pixel-edge): accumulate content before finish reason in stream fallback guard

[vaibhavsrv]

Why this matters

In the baseline Pixel Edge streaming proxy (_stream_upstream), concise model completions or single-chunk model responses (common with smaller, faster models or inference engines emitting the final token and finish_reason: "stop" concurrently in the same chunk) trigger the empty-reply fallback prematurely. Because finish_reason was evaluated before accumulating content into pending_text, pending_text was evaluated as empty "", causing the edge to inject empty_reply_fallback ("I couldn't produce a useful response to that request...") before the actual model content. Furthermore, when replace_pending was called on unverified turns, the unparsed line was written unconditionally afterwards, leaking reserved phrases like NO_REPLY into the client stream.

This change accumulates incoming SSE content into pending_text before evaluating finish_reason, and cleanly synthesizes finish_reason: "stop" when replacing pending empty or reserved replies (synthesize_finish=True) without leaking the unverified or reserved payload. Existing SSE chunk framing limits, CRLF normalization, and reserved reply definitions remain completely untouched.

Validation

• Baseline reproduction: On unpatched code, a model response containing {"choices": [{"delta": {"content": "Paris"}, "finish_reason": "stop"}]} failed with 'FALLBACK_CALLED' unexpectedly found in the stream output, erroneously prepending the fallback error message before "Paris". Single-chunk NO_REPLY responses leaked 'NO_REPLY' into the output after the fallback message.
• Post-fix verification: The single-chunk response delivers "Paris" cleanly without triggering the fallback; reserved NO_REPLY payloads are cleanly replaced with synthesized fallback frames without leaking NO_REPLY. The full test suite in test_pixel_edge.py (102 tests) and new regression test_pixel_edge_single_chunk.py (3 tests) pass cleanly.
• Telemetry statement: "Pixel Edge suites: 105 passed. New-test Ruff, PyCompile, and diff checks pass; new regressions wired into Linux CI."

Overlap check

• PR fix(pixel): guard task group exception groups during stream revocation cleanup #5600: Guards task group exception groups during stream revocation cleanup in pixel-inference.
• PR fix(dashboard-api): coerce null metric values in resource totals calculation #5469: Hardens metric value null guards in dashboard-api resource calculation.
• PR Fix public-beta fleet install and Pixel runtime blockers #4215: Resolves fleet installer and runtime preflight blockers.
  This PR touches only pixel_edge.py SSE event accumulation and does not overlap with or invalidate any of the above changes.

Risk / AI disclosure

AI-assisted investigation, implementation and CLI regressions. This strengthens an SSE stream delivery check, not runtime admission or model authorization. Independent human review and platform/runtime qualification remain gates. No running configuration, deployment or upstream merge changed.

Follow-up integration evidence

Composed with #5600 at 3f44dce without conflicts. Production and test diffs passed together; pixel edge and inference checks remain intact.
Backlog composition was local-only (production/test diffs, excluding workflow/Makefile wiring); it is not an upstream merge or independent human approval. Declared live-review gates remain open.