fix(agent): mid-turn compact gate fires on the server meter, not just the local estimate

[justrach]

Problem

A long-running codex/.responses turn died mid-turn: after a tool call returned, the next request was rejected with "Your input exceeds the context window." graff auto-recovered (compact ~511k → retry → summary), but the #193 pre-send gate should have compacted before shipping the over-cap request. It didn't.

Root cause — two gates, two different signals

• Between-turns gate (mainloop.zig:683/734) trusts last_context_tokens, the accurate server-reported count — but only runs after a turn finishes.
• Mid-turn pre-send gate (inputOverCompactThreshold) trusted only the local fullInputEstimateTokens byte/4 estimate, and runs every loop iteration.

The local estimate structurally under-counts codex histories: retained encrypted-reasoning items are billed large server-side but are compact base64 handles locally. So during a long single turn the local estimate stayed under the threshold while the server had already reported ~511122 via last_context_tokens (refreshed mid-turn by recordUsageResponses) — and the mid-turn gate simply never looked at that number. It had the right answer in hand and ignored it.

Fix

Make the mid-turn gate also fire on the server meter — the exact signal the between-turns gates already trust:

if (self.last_context_tokens >= threshold) return true;

Plus two supporting changes:

• Test correctness: the #193 test builds Agent = undefined, so the new last_context_tokens read was uninitialized-memory UB that would deterministically break the suite. Explicitly zero it, and add a meter-only case that trips the gate with an empty local history.
• Hardening: fullInputEstimateTokens now falls back to the bytes counted so far on a serialize failure (catch return d.fullCount()/4) instead of a spurious 0, so a failed serialize can't blind the pre-send gate and the #202 meter floor at once. (Zero-risk: the Discarding sink is infallible today; this future-proofs a stricter sink.)

Residual gap (unchanged from prior behavior)

A cold-start giant first request of a turn, while last_context_tokens is stale-low from a prior small turn, is still guarded only by the local estimate + the #201/#203 output caps. The reported failure mode (multi-round-trip, meter already at 511k) is now caught.

Verification

• zig build clean
• zig build test → 195/195 unit tests pass (incl. the modified #193 test exercising the new meter path)
• zig build repl-test → 19/19 REPL tests pass
• Diff: +21/−1 in src/agent_request.zig