fix(agent): context-overflow follow-up — recovery-path parity + local-model wedge (#201-#204)

[justrach]

Follow-up to #196, closing the ranked backlog from the context-overflow audit.
Together these break the compound local-model wedge chain end to end: bad window
guess -> undercount -> gate never fires -> needle miss -> no recovery -> hard wedge
(previously escapable only by a manual /compact).

Stacked on #196 (base = fix/193-tool-output-cap); GitHub will retarget this to
main when #196 merges, at which point the Closes keywords fire.

Commits

• ea230cc — G1/G3/G6/G10 (context-overflow: perOutputCap × keep_recent=4 lets recent tool outputs wedge the window (hole in #193 per-output cap) #201, context-overflow recovery hardening: usage char/4 fallback + visible truncation + meter re-estimate (anthropic/openai parity) #202): the recovery-path holes
• 3a603d0 — G8/G4/G2/G5/G9 (local/unknown models wedge on context overflow: 200k default window + system-prompt-blind estimate + substring-only detection #203, context metering parity: same-format provider switch carries stale token count; compaction threshold hardcoded at 80% #204): local-model window + metering parity

What changed

Gap	Fix
G1 #201	perOutputCap was ~50% of the window each; with keep_recent=4 verbatim, four recent outputs pinned ~2x the window past recovery. Now ~1/8 window + 256 KB ceiling, so 4 recent stay reclaimable.
G3 #202	recordUsage froze the meter on missing usage; now floors at max(fullInputEstimateTokens, body/4) (the fallback the codex path already had).
G6 #202	capOversizedToolOutputs truncated silently; now emits a tracer note + status line.
G10 #202	in-place trims zeroed the meter; now re-estimate the trimmed size.
G8 #203	unknown/local models defaulted to a 200k guess. GRAFF_CONTEXT declares the real window, applied only to non-catalogued models (pricing.isKnownModel).
G4 #203	the pre-send gate estimate omitted system+tools; adds a window-clamped baseline (kept out of fullInputEstimateTokens, which stays pure).
G2 #203	overflow detected only by English substrings; now matches structured error.code first (codex parity), substrings as fallback.
G5 #204	a same-format provider switch kept the old model's token count; now re-estimates.
G9 #204	80% compaction threshold hardcoded; GRAFF_COMPACT_PCT overrides it (both directions).

Design note: the openai/codex research behind the local-model decisions (window
override vs probe vs catalog, structured detection) is captured in #203.

Tests

zig build test 183/183 (was 179; +perOutputCap, +recordUsage-fallback,
+contextWindowFor, +compactAt-override; overflow-detection tests gain structured-code
cases), zig fmt clean.

Closes #201
Closes #202
Closes #203
Closes #204

[justrach]

Added an end-to-end overflow PTY test (scripts/test-pty-overflow.py) that points graff at a mock lmstudio backend injecting a context_length_exceeded error and drives a real terminal turn.

Writing it caught a real gap in the G2 fix: the structured-error.code detection only covered the anthropic-style {"type":"error"} envelopes. OpenAI-compatible errors are {"error":{...}} (no top-level "type") and land on the generic apiErrorMessage path, which was still passing null for the code — so a local provider (LM Studio / mlx / any openai-compat base_url) whose message text we don't match on would not have been recognized as an overflow. That's the exact case G2 exists for. Fixed with errorCode(root) (aa2b995).

The test proves both directions:

• code=context_length_exceeded + a Dutch message (matches no English needle) → meter pins to 200k/200k (100%) and the REPL stays responsive → detection went through the structured code, no wedge.
• code=rate_limit_exceeded + non-overflow message → meter must not pin → detection is precise.

Deterministic across repeated runs (SO_REUSEADDR on the port probe); skips cleanly if 127.0.0.1:1234 is already held by a real LM Studio.