The Synthesis Lost Detail the Panel Actually Gave Us

The Broader Goal at the End of the Previous PR

PR #1919 synthesized three synlynk decide panel runs into a consolidated architecture roadmap, a surface-simplification plan, and a broader-issues remediation plan — the direct follow-through on the five-POV review's open questions. The architecture roadmap compressed nine distinct improvement opportunities from the review down to seven table rows.

Strategic Shifts in This PR

No code changed here either, and the roadmap's substance doesn't change. What changed is fidelity: checking the shipped table against the original nine-item list showed the synthesis pass had been too aggressive. Three items — a HarnessAdapter plugin registry for third-party harnesses, a CI release-version gate, and an explicit cold-start performance budget — were dropped entirely rather than reconciled or deferred. Two more — the dispatch_agent pipeline's actual stage shape, and containerized execution for untrusted harnesses — survived only as a passing mention, missing the specifics the panel's own per-harness responses had given. This is a real lesson about this kind of synthesis step: compressing multiple detailed inputs into a short table is exactly where information gets silently discarded, and the discard isn't visible unless someone checks the table against the raw inputs it claims to summarize.

What Shipped

• docs/strategy/2026-10-02-decide-panel-roadmap.md, revised: the architecture roadmap table grew from 7 rows to 10, recovered from the session's captured per-harness panel output rather than re-running the panel (the original responses were still available).
  • Row 5 now names the actual pipeline shape (resolve → authorize → prepare_worktree → spawn → observe → finalize) and the full HarnessAdapter method set (build_cmd(), parse_events(), permissions(), detect_quota(), models()), plus the plugin-registry detail for third-party harnesses.
  • A new row 6, safe-by-default execution, separates the permission-default flag flip (already covered, in the broader-issues plan) from containerized execution for untrusted harnesses (not covered anywhere until now), and cross-references both documents so the shared flag change lands once, not twice.
  • Row 8 (CLI core/packs split) now names the actual ~15 core commands and the 15–20K LOC target instead of a bare direction.
  • New rows 9 and 10 restore the CI release gate (tie to the live version-drift issue, #1914) and the lazy-import/cold-start performance budget, both previously missing entirely.
  • The sequencing decision and the cross-cutting note were updated to match the new item numbers.

What This Achieved on the Path to Autonomy

The original panel-decide loop worked end to end — real multi-harness deliberation, real synthesis, real committed output. This PR closes the gap between what the panel actually said and what the committed document claims it said, which matters more for a process meant to be trusted as a planning input than any individual roadmap item does.

Strategic Note: The Goal at the End of This PR

The architecture roadmap and both companion plans are now a complete, faithful representation of the five-POV review's findings. The next milestone is unchanged from before this patch: ship roadmap items #1+#2 (state/worktree GC and the regen-bug guard) first, since they're small, parallel-safe, and stop data loss that has already recurred multiple times.