fix(agent): bound local_loop config and stop returning dead stack storage - #1046

Third and last PR splitting #987. Addresses the boundary findings. #1044 was the gating fix, #1045 the concurrency fix.

1. Slice into a dead stack frame

fn stackLowerAscii(input: []const u8) []const u8 {
    var buf: [256]u8 = undefined;
    ...
    return buf[0..cap];        // ← dead on return
}

extractErrorSignature then ran indexOf over that invalid slice. The helper is deleted and the scan is case-insensitive in place, so nothing escapes the frame. Detection behaviour is unchanged — there's a mixed-case test to prove it.

2. Truncation marker exceeded the budget

max_result_chars: 1 returned the whole 16-byte "\n… [truncated]": the code computed body_len = cap -| suffix.len, got 0, and emitted the suffix anyway. When the budget can't hold the marker, output is now a literal truncation with no marker — the cap is never exceeded. Tested across budgets 0–16.

3. Explicit max_result_chars was indistinguishable from unset

The field defaulted to 8192 — the same value the documentation's own example uses. So setting it explicitly to 8192 could not be distinguished from omitting it, and the 400-character tightening cap silently replaced it. That directly contradicts:

local_loop.enabled = true tightens the default tool-result history cap to 400 characters (unless max_result_chars is set explicitly).

0 is now the "unset" sentinel, resolved per mode in toolResultCompressOptions: the tightening cap applies when the feature is on, the general default otherwise. I resolved it there rather than in the struct default deliberately — that keeps the sentinel correct on its own, independent of whether compression is separately gated on enabled (which is #1044's change). Stacking that assumption would have made this PR secretly depend on the other.

4. Unchecked casts

All six numeric local_loop fields were @intCasted unchecked, so a positive value wider than u32 trapped in a safety build instead of reporting a configuration error. Each now returns error.InvalidLocalLoopConfig.

One boundary worth recording: a literal beyond i64 (e.g. 1e30) never becomes an integer token, so the JSON parser drops the field rather than the bounds check rejecting it. The tests use 5000000000 — wider than u32, still representable as i64. Tightening the tokenizer is a separate concern and is not attempted here.

Tests

• budgets 0–16 never produce output longer than the budget
• error-signature detection stays case-insensitive (mixed case, upper case, and a no-marker line)
• all six fields reject a value wider than u32; in-range values still parse
• explicit 8192 is honoured; unset applies the 400 cap

Validation

• Branch: zig fmt --check src/ exit 0 · ReleaseSmall exit 0 · zig build test --summary all 13/13 steps, 7400/7409 passed, 9 skipped, 0 failures, 0 leaks
• Merge result with current main: clean merge, fmt exit 0, ReleaseSmall exit 0, 13/13 steps, 7518/7527 passed, 9 skipped, 0 failures, 0 leaks

Docs

Both languages updated: the example now shows max_result_chars: 0 and the note explains that 0 means unset and any other value is used as-is.

#987 split — complete

With this open, #987 can be closed with a finding-to-PR mapping and its authorship preserved. All three build on feat/local-loop-phase0, since that is where the affected code is introduced — none of them have a target on main until the feature branch lands.