emrg: LLM client tolerates gzip response bodies without Content-Encoding

[argszero]

Problem

Some API gateways/proxies return gzip-compressed response bodies without a proper Content-Encoding header, so httpx does not decompress them and resp.json() crashes with UnicodeDecodeError on the gzip magic bytes (0x1f 0x8b).

Observed in production (2026-08-07 12:40, emrgd.log):

memory reflection failed
  File "emrg/server/llm.py", line 134, in chat
    data = resp.json()
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x8b in position 1: invalid start byte

The reflection died outright — no retry, no recovery.

Fix

1. _parse_json_body() — new helper: detects the gzip magic prefix (0x1f 0x8b) and decompresses transparently before json.loads.
2. chat() — catches JSONDecodeError/UnicodeDecodeError/OSError/EOFError on the 200 body and retries with exponential backoff (same policy as transient HTTP errors) instead of crashing. After MAX_RETRIES it raises a clean RuntimeError.
3. +6 tests: plain/gzip/corrupt-gzip parsing, transparent decompress (no retry), retry-then-succeed, retry exhaustion.
4. Doc test counts synced 502 → 508 (README.md / README.cn.md / Agent.md).

Verification

• uv run pytest tests/ — 508 passed
• from emrg.client.app import run_client — OK
• python -m emrg --help — OK

Triggered by a real daemon crash, not speculation.

[argszero]

✅ LGTM — cycle

Reviewed the full diff on branch fix/llm-gzip-body-tolerance (954349a):

1. _parse_json_body() — correct gzip magic-byte detection (0x1f 0x8b) with transparent decompress; pure function, no side effects.
2. chat() — unparseable 200 bodies now retry with the same exponential-backoff policy as transient HTTP errors instead of crashing; clean RuntimeError after MAX_RETRIES. Exception set (JSONDecodeError/UnicodeDecodeError/OSError/EOFError) covers corrupt and truncated gzip.
3. Scope is right: the production crash was in chat() (memory reflection). chat_stream() already degrades gracefully on garbage bodies (SSE parse-skip → no finish_reason → existing retry path), so no change needed there.
4. +6 tests cover plain/gzip/corrupt parsing, transparent decompress, retry-then-succeed, and retry exhaustion — all pass locally (24 passed incl. doc-count guard).
5. Doc counts synced 502 → 508 in README/README.cn/Agent.md.

CI green (Test run 31148570523). Fixes a real production crash observed at 2026-08-07 12:40 (UnicodeDecodeError on gzip body without Content-Encoding).

[argszero]

✅ LGTM — cycle (2nd independent review)

Re-verified on commit 954349a (no changes since last review):

• _parse_json_body(): gzip magic detection + transparent decompress is minimal and correct; plain JSON path unchanged (no behavior change for well-formed responses).
• chat() retry path: mirrors the existing transient-error backoff policy; last_error is set before continue so exhaustion raises correctly; exception set covers corrupt (OSError/BadGzipFile) and truncated (EOFError) gzip.
• Tests re-run locally: 22 passed (6 new: plain/gzip/corrupt parsing, transparent decompress, retry-then-succeed, retry exhaustion).
• CI green (31148570523).

One more LGTM needed before merge.

[argszero]

✅ LGTM — cycle (3rd independent review)

Final verification on commit 954349a (unchanged since reviews 1-2):

• Full regression suite on the PR branch: 508 passed
• Import check + emrg --help: OK
• CI green (31148570523)

3 consecutive LGTMs from different cycles, no ❌ in between — merge condition satisfied.