Hand off minimal decision routing through its instruction source owner - #571

Scope

On a host whose personal directives are generated by another source owner, the previous full-pack helper appended another 56/62-line defaults/RTK pack. Export only the existing three-line decision-routing fragment, refuse generated personal output, and preview only an instruction source the caller already owns. decision-md never writes, backs up or replaces a target; the producer owns mutation.

• Base commit: 5597f9fae2c6148037a74b4e73d2c55fc11c9ff6
• Lane: lane:foundation
• Owned paths: adoption/templates/decision-routing.md, adoption/update.md, the dated decision and one anti-pattern row in docs/, tools/adoption/managed_block.py, the explicit caller in adoption/bootstrap-linux.sh, and regression coverage in tests/test_managed_block.py / tests/test_bootstrap_full_profile.py; manifests/evidence.json only registers their hashes in the last commit.

The Mac handoff exports the fragment to the separate agent-ecosystem producer owner. This PR neither edits that producer nor installs directives/catalog data on the Mac. Its reviewed producer has no selective installed-registry command: broad catalog_refresh.py --apply invokes the full installer and preserves decision/import pins. The producer owner must reconcile maintained selections/retired-host routes and qualify a receipt-aware selective synchronization path there. No replacement registry is synthesized.

SOTA sources

• ECC search-first at c70874fa: coordinator read and SHA-256 verified against the pinned manifest; extend the existing integration instead of adding a router/installer.
• Codex rust-v0.159.3 instruction loader: preserve override order and native whitespace behavior.
• Claude Code v2.1.286 changelog and native memory contract: retain the existing Markdown integration. Current memory-page fetch was denied; no fresh fetch is claimed.
• agent-ecosystem directives.py at 753777ac, install_local.py, and catalog_refresh.py: verified source ownership, generated notice, broad installation scope and receipt hashes.
• Existing managed_block.py supplies native Markdown marker parsing and preview behavior. apply_claude_settings.py is unchanged from the base; the new writer was removed because a check followed by replacement cannot guarantee atomic compare-and-swap. Bootstrap passes its existing home-scoped Claude target explicitly. The fragment is checked against the unchanged canonical Codex template paragraph.

Evidence-class table

Local commands run

python3 -m unittest -v tests.test_managed_block tests.test_apply_claude_settings tests.test_adoption_docs_consistency tests.test_bootstrap_full_profile
# Final export/preview and caller controls: 127 run, 126 passed, 1 existing skip; exit 0.
# Earlier overlapping suites and their diagnosed failures are retained, not summed.
python3 scripts/component_matrix.py --write
python3 scripts/new_host_grand_list.py --write
python3 scripts/validate.py
python3 scripts/validate_catalogs.py
python3 scripts/build_ecosystem.py --write
python3 scripts/build_ecosystem.py --check
git diff --check
# All exit 0; generated reports unchanged, own hashes registered through host_receipts.register_file.

Decision record

docs/decisions/2026-10-01-minimal-owned-source-sync.md records the reported duplication, alternatives, producer boundary, citations and conditions that reopen this choice. Original local failed fixture/doc checks were diagnosed and retained in the cloud logs; assertions were not weakened.

Independent read-only reviews covered the producer interfaces, evidence scope, instruction formats and write safety. Final read-only review found no unresolved material blocker after addressing the three GitHub review findings; source commit c1a6585dc7d859da7cc36771d2ee67062084fa22 binds the reviewed content. Export/preview only removes the unsafe writer; normalized paragraphs detect wrapped copies; the integrated bootstrap caller supplies its intended target. Earlier passed checks at 0071c67c do not qualify the corrected head.

The reported Mac production remains ai-memory 2.5.0 / Ollama 0.34.4; Hindsight 0.10.2 remains isolated. The 20-pass/2-failure lifecycle and the 20-second mismatch remain open despite the later two parsed-JSON diagnostic calls. Measured and blind promotion gates, native caching/compaction, models/effort, QMD relevance failure and surviving-host preregistration remain intact.

Host evidence

No files under evidence/hosts/ changed. The sanitized owner handoff is reported host evidence, not execution by this cloud task.

Checklist

• No workflows changed; existing pinned actions and permissions retained.
• No secrets printed, logged or committed; no new required secret.
• No new paid hosting, subscription or billing surface.
• Peer-owned untracked files, source repositories and host globals preserved.