Enhancement: Promote adaptive planning and checkpointed operator switching

[smiggleworth]

Scope

Establish the evidence and controls required before adaptive planning or checkpointed operator switching can move beyond Experimental/opt-in status.

Source of truth

docs/product-roadmap.md:27-31 and docs/experimental-promotion-criteria.md:20-33.

Acceptance criteria

• Define representative deployment profiles, confidence/cost thresholds, stale-feedback behavior, and rollback controls.
• Prove result equivalence and deterministic behavior against the fixed-plan baseline.
• Add EXPLAIN/metrics guard diagnostics and failure/fallback evidence.
• Keep the feature disabled by default until all gates pass.
• Update support/readiness status only for the selected surface.

Tracking: #1.

[smiggleworth]

Full-history re-audit at current main d0874ff: this issue has no closing implementation or acceptance comment; docs/feature-support.md still marks adaptive planning Experimental and docs/product-roadmap.md still requires it to remain opt-in until representative evidence exists. Remaining criteria: representative profile and threshold contracts, fixed-plan equivalence, deterministic switching, diagnostics/fallback evidence, and an explicit support-status decision. Next action: audit existing adaptive-planner tests and benchmarks, then implement the first missing evidence slice without enabling the feature by default.

[smiggleworth]

Current-main re-audit at 8b84ee936986f586567530a2decee949d2b8f14c: the repository now defines two explicit evaluation profiles in docs/experimental-promotion-criteria.md. disabled-default is the rollback/default state; evaluation-conservative names confidence and cost thresholds. Current adaptive tests cover fixed-path defaults, result-shape equivalence, stale/low-confidence guards, deterministic EXPLAIN fields, metrics, and vectorized-to-merge operator switching/fallback behavior.

The promotion gate remains incomplete. The evaluation profile is not backed by retained representative same-commit benchmark/deployment evidence demonstrating that its 500-basis-point savings and 900-basis-point confidence thresholds are appropriate, and no production profile supports enabling adaptive execution by default. Canonical run 32505662069 is incomplete; PR #158 addresses the immediate Tier 3 blocker but has not yet produced a complete artifact.

Disposition: keep #12 open, Adaptive planning Experimental, and both adaptive execution and operator switching disabled by default. Dependency: #7 plus representative fixed-versus-adaptive evidence under a named deployment profile. Before promotion, inspect result equivalence, selected operator, confidence/cost guard diagnostics, stale-feedback fallback, worker/resource cleanup, trust/noise, and rollback to disabled-default; do not tune thresholds from a failed or diagnostic-only run.