PR #2082 — Empirical routing promotes a harness only after five measured samples

The Broader Goal at the End of the Previous PR

v0.24.0 still needs dispatch to choose a harness from measured outcomes, and to say so honestly in the README. PR #2074 made a missing GOVERNS link visible before #1990 can hard-fail on it. Routing itself was still the suspended heuristic table.

Strategic Shifts in This PR (if any)

Single-workspace routing no longer waits on shard consolidation (#1926). Each repo's dispatch preview reads that workspace's own canonical state.db. Fleet-wide synlynk capability report aggregation (#1993) stays blocked on #1926 and #1831. That split is what lets empirical promotion ship now.

What This PR Shipped

• ranked_harness_for_task() in synlynk/capability.py. candidates[0] is the incumbent. A later harness is preferred only with at least five capability_ratings rows whose discipline matches the task type (there is no task_type column) and a joined cost_entries row on story_id where ce.harness equals cr.agent. It must beat the incumbent on both median pr_review_cycles and average total_cost_usd. Otherwise the function returns None.
• compose_dispatch_preview() is the public name for _infer_dispatch_defaults() at the task_allocation read. When no explicit agent, role, or story is bound, a promoted harness replaces the table default. Explicit CLI values and story metadata still win. The local-threshold path in _resolve_dispatch_agent() is unchanged.
• .synlynk/policy.json capability_policy drops suspended_since and suspended_rationale, restates blocking_dependency for the single-workspace vs fleet split, and names README.md:15 next to task_allocation and docs/harness-capability-baseline.md for any future suspension.
• README line 15 now says routing uses the best-measured harness once it clears the sample-size bar, and the documented default otherwise.

Tests seed an in-memory ledger (creating cost_entries, which lives in the db.py migration rather than _DB_SCHEMA) and a dispatch preview that stubs the ranker. tests/test_governs_linkage_warning.py still passes.

Plan: docs/superpowers/plans/2026-10-06-three-lens-followups-plan.md Task A. Design: docs/superpowers/specs/2026-10-06-three-lens-followups-design.md Section A. Issue: #2063.

Brainstorm Visuals Used

None. This follows the committed design and plan.

What This Achieved on the Path to Autonomy

Dispatch preview can now prefer a harness the workspace has actually measured, and it refuses to do that on a thin sample. Below five rows the documented allocation stays in force, which is the honest default until grok and muse are calibrated.

Strategic Note: The Goal at the End of This PR

The next follow-up is durable logging of routing fallback (#2064): the decision this preview makes has to survive in sentinel.md and job metadata, not only on stdout. Fleet-wide capability reporting remains behind #1926.