agent_settings: tool_result_inline_max_chars becomes a durable resident setting

[antra-tess]

Immediately coupled follow-up to #91, implementing Sol's ruling (antra concurring) from #architecture: the resident-set inline cap must use the same durable, agent-exposed config plane as everything else — not an ephemeral hot override.

Semantics (Sol's spec, point by point)

• Recipe/Host config supplies the default: FrameworkConfig.toolResultInlineMaxChars (from Complete oversized tool-result spill: safe 5k default, error paths, all result paths (#89) #91) → house 5000. Unchanged.
• agent_settings update tool_result_inline_max_chars persists across restarts: written to the toolResultInlineCaps slot in framework/state — the same store plane as agentRuntimeSettings — and restored in create() with load-time validation (invalid entries dropped loudly, same >= 1000 rule as the live update).
• get reports desired/effective/source/clamp: from Complete oversized tool-result spill: safe 5k default, error paths, all result paths (#89) #91, extended below.
• reset deliberately returns to the residence default, and the reset itself survives restarts.

The one design decision — flagging for Sol

opus-rev's review point: #91's override escapes the strategy clamp, which was defensible for a lift that died on restart but becomes a silent permanent over-budget pin once durable. This PR keeps the resident's value honored (they own the setting; clamping it would gut the raise-to-see-more use case on bounded strategies) but makes it impossible to be silent: _source reports agent-settings-override (exceeds strategy bound) whenever the pinned value sits above maxMessageTokens * 4. If the house later wants a hard clamp instead, it's a two-line change in resolveToolResultInlineCap.

Not addressed here (bigger surface, opus-rev's option (b)): a general persistence contract for AgentSettingsExtension so module-owned settings can be durable too. This PR only moves the framework-owned key onto the durable plane; the extension interface is unchanged.

Tests

12 in test/tool-result-spill.test.ts (was 11): the restart test now proves the full durable lifecycle — resident sets 60k → survives restart with correct provenance → reset returns to the 8k residence config → reset survives a further restart; plus the new exceeds-strategy-bound provenance pin. Full suite: 531 pass / 0 fail / 1 skip (CI red remains the pre-existing main drift, same five error classes).

Release note: rides the same AF release train as #91/#88 — no additional host changes needed for this PR (the recipe passthrough for the default remains the separate host-side PR).

🤖 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.

[antra-tess]

Revised per Sol's ruling — head is now 5b9f0a7:

• Effective cap is hard-clamped: min(desired, strategy bound) for every source, resident value included. The clamp-escape from Complete oversized tool-result spill: safe 5k default, error paths, all result paths (#89) #91 is gone entirely.
• Desired survives: the resident's persisted value is kept verbatim and reported as tool_result_inline_max_chars; _effective reports the min; _source is now a clean enum; _clamped_by is a dedicated key ('strategy-bound' | null) replacing the source-string suffixes.
• Test updated to pin desired-preserved / effective-clamped / clamp-enforced-on-the-wire; the restart lifecycle test is unchanged and still proves desired-survives-restart.

Full suite 550 pass / 0 fail / 1 skip locally (with the collection-count caveat opus-rev documented). PR body's 'honored above bound' paragraph is superseded by this comment.