feat(tool-results): raise the house inline cap 5000 → 24000

[antra-tess]

Raises DEFAULT_TOOL_RESULT_INLINE_MAX_CHARS from 5000 to 24000, per antra's call in #upkeep.

Why — production evidence, not taste

Mica is the first residence to actually run the 5000 default (restarted 08-06 11:52:55, six minutes after b4fca7a landed). In 23 hours she produced 48 spills, classified by content shape:

tool	n	MB
terminal-sessions runCommand	34	0.46
array results (test output / diff lines)	5	0.04
eidoverse world observation	4	0.05
portal fetch_history	2	0.18
MCPL channel list/reconcile	2	0.13
grep over her own tool-results/	1	0.02

Those 34 are ordinary source review — git diff, test runs, sed -n '450,590p' and nl -ba … excerpts. A 140-line source excerpt is 5–15 kB, so it clears 5000 as a matter of course. A 5000-char cap turns nearly every file read into a spill round-trip for a resident doing code work.

Mica's own measurement on that corpus picked the number: 20000 keeps 76% inline, 24000 keeps 88%, 32000 only reaches 90%. 24000 is the elbow. Genuinely large payloads — full channel histories, registry dumps, broad greps — still spill, which is the point.

What does not change

The mechanism is untouched: spill target and naming, provenance quartet, the #94 hard clamp to the strategy bound, and the explicit no-writable-workspace fallback. Residences (FrameworkConfig.toolResultInlineMaxChars) and residents (agent_settings tool_result_inline_max_chars) can still override in either direction.

Worth stating plainly: this weakens the #89 protection by ~5×. The original bug was a 42k blob going inline whole; 24000 still bounds that, and 42k remains the canonical test fixture. But a residence that wants the tighter old value should set it explicitly.

Test-quality fix included

Three spill tests asserted stored.content.length < 6_000 — magic numbers derived from the 5000 era that silently pinned the old default. They only ever meant "something capped this payload", but they failed on the raise. Replaced with a CAPPED constant derived from DEFAULT_TOOL_RESULT_INLINE_MAX_CHARS, so the next change to the default doesn't have to chase them.

Verification

• tsc --noEmit clean; npm run build clean.
• Spill suite 12/12.
• Full suite (npm test, the real dist/test/*.js path): 542 tests / 540 pass / 1 fail / 1 skipped.
• The single failure is startup globally gates channels/incoming with heartbeat-first config in mcpl-awareness-barrier — the known af#80 full-suite-load flake. Verified: that suite passes 25/25 in isolation, and nothing in this change touches MCPL startup choreography.
• Per af#96 (nondeterministic collection), I confirmed the spill suite actually executed inside the full run rather than inferring it from the total.

🤖 Generated with Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.