compaction: allow an explicitly configured compaction model/provider with safe fallback

## Problem

Compaction summary generation always uses the session model and its auth (`_getCompactionRequestAuth` -> session model). When the session model is native Cursor and the conversation is already under context pressure, the compaction request itself can fail with `resource_exhausted` (observed in session 01a01831: compaction tokensBefore ~4.2M, cursor/claude-fable-5-high, two compacts saved ~1% before the turn died).

PR # fixes the surrounding death spiral (threshold now samples the local transcript estimate, token-bearing `resource_exhausted` classifies as overflow, billed usage lands on `usage`), so compaction should now fire long before the context is hopeless. This issue tracks the remaining hardening.

## Proposal

Allow an explicitly configured compaction model/provider with a safe fallback:

- New optional compaction setting, e.g. `compaction.model: "provider/model"`, resolved through the normal model runtime + auth path.
- When configured and resolvable, compaction requests use that model; otherwise fall back to the session model (current behavior).
- No silent automatic provider switching: changing provider mid-compaction affects credentials, cost, and data-routing expectations, so only an explicit configuration may route compaction elsewhere, and a clear failure must surface when the configured model cannot be resolved.

## Notes

- Reviewer guidance from the report thread: prefer explicit configuration over an automatic cheaper-provider heuristic.
- Needs: settings schema + docs, auth resolution reuse, availability check, tests for configured/fallback/unresolvable cases.

Start with the compaction request path around _getCompactionRequestAuth, then trace the settings schema and normal model runtime/auth resolution. Check existing compaction and authentication tests before adding coverage for configured, fallback, and unresolvable cases. Done means the schema, documentation, availability behavior, and all three model-selection paths are covered.

Written by the indexing model from the issue text.