git:20260706.367aa8b to git:20260915.64f0bee

+ | `model` | string (optional) | model for EVERY spawned agent (`agentOpts`), overriding each agent type's own default (e.g. `lets:explorer`'s `sonnet` pin). Must match `^[A-Za-z0-9.\[\]_-]{1,64}$`, else ignored. Absent = each agent's default. |

+ | `budget` | `{explorers, approaches, judges, evaluators, plan_reviewers, plan_checker}` (optional, any subset) | per-stage caps that can only LOWER a stage (`capStage` keeps the first N, logs the rest). Each value is clamped to MIN..MAX (`explorers`/`approaches`/`judges` 1..5/4/5, `evaluators` 0..5, `plan_reviewers` 0..4, `plan_checker` 0..1); a clamp is logged. `approaches` = how many proposed approaches get architected and judged (the Approaches prompt then asks for best-first order). `plan_reviewers: 0` / `plan_checker: 0` skip that pass as a named skip, never as a failure. Send only the stages the user changed. |

+ `{ plan_markdown, delivered_approach, diverged_from_winner, divergence_reason, review_findings[], check_findings[], refinement_log{}, decision_log, winner, winner_name, approaches[], eval_findings[], counts{} }`. `plan_markdown` is the FINAL self-repaired plan (after Plan Review -> Revise -> Plan Check -> Refine). `refinement_log` = `{review_verdict, review_findings, review_fixed, review_failed, review_skipped, check_verdict, check_findings, check_fixed, check_failed}` - the self-repair audit (if reviewers/checker all errored, the prior plan is kept and the `*_failed` flag says so; never silently skipped). In fast mode the heavy review pass is skipped, so `refinement_log` carries `review_skipped: true` with `review_verdict: null` and `review_failed: false` (a deliberate fast-skip is NOT a failure), while `check_verdict`/`check_findings`/`check_fixed`/`check_failed` carry real values in BOTH modes (Plan Check always runs). `counts.mode` is `'fast' | 'standard'` and lives INSIDE the `counts` object on EVERY return - both the success return and every typed early-error return (`exploration_failed`/`no_approaches`/`architecture_failed`/`judge_failed`/`plan_synthesis_failed`) - so a failed fast run is distinguishable from a failed standard run. `decision_log.forced = 'single-candidate'` marks the fast judge-fallback. `counts.budget = {model, caps, drops[]}` is also on EVERY return; a run without `model`/`budget` gets `{model: null, caps: {}, drops: []}` and is otherwise unchanged. A drop is `{stage, kept, dropped, dropped_labels[]}` (a capped list) or `{stage, clamped: true, requested, applied}` (an out-of-range value). A budget skip adds `refinement_log.review_skip_reason: 'budget'` (with `review_skipped: true`) or `check_skipped: true, check_skip_reason: 'budget'`; the keys are absent otherwise. Anti-silent-fail: each upstream stage that wipes out (explorers / approaches / architects / judges all errored, or plan synth returns null) returns an `error` + null `plan_markdown` instead of fabricating - the dispatcher surfaces it and offers a re-run. `decision_log` = `{votes, totals, rationales, judges}`.