PR #2083 — Routing fallbacks are written down, not only printed

The Broader Goal at the End of the Previous PR

v0.24.0 still needs a dispatch to move from implementation to a trusted review verdict on GitHub state alone. The previous PR made a missing GOVERNS link visible in the log and in sentinel.md, while #1990's hard-fail stays unbuilt. The next visibility gap in the same class is auto-routing: when local cannot run, dispatch already prints the fallback and then forgets it.

Strategic Shifts in This PR (if any)

No strategy change. Issue #2064's title talks about failing closed. The committed design keeps today's routing and only records the decision. Fail-closed stays out of this PR.

What This PR Shipped

• _resolve_dispatch_agent() still prints Routing to: <fallback> when local oMLX is unreachable or the local capability score is under the threshold. The same line is appended to .synlynk/sentinel.md as DISPATCH_ROUTING_FALLBACK through the existing _write_sentinel_alert helper.
• The job dict and daemon_jobs gain requested_harness, actual_harness, and fallback_reason. A fallback from local to codex is stored as requested local, actual codex.
• synlynk jobs prints local->codex on that row. A dry run discards the pending decision so it cannot attach to a later job.
• Choosing local, or naming a harness explicitly, does not write a fallback alert. Return values and control flow are unchanged.

Plan: docs/superpowers/plans/2026-10-06-three-lens-followups-plan.md, Task B. Design: docs/superpowers/specs/2026-10-06-three-lens-followups-design.md, Section B.

Brainstorm Visuals Used

None. This follows the committed design and plan.

What This Achieved on the Path to Autonomy

Job history can show that a run asked for local and executed the fallback. Cost attribution and synlynk jobs can read that from the job row instead of inferring it from a harness name that was already substituted.

Strategic Note: The Goal at the End of This PR

The routing decision is now durable. The next piece of this follow-up set is still the sample-size-gated empirical ranking (#2063), the config schema checks (#2065), and the adapter conformance suite (#2062). #1990's GOVERNS hard-fail remains the enforcement step for the warning shipped just before this.