docs: screen candidate runtime-worker context policy - #1181

Scope

Screen a candidate context policy for a future TRIAL on Pi, OpenCode, fenced Muse and Claude Code workers: retain vendor compaction thresholds, propose a bounded current-state note and explicit handoffs at task boundaries. Record one unrun comparison of native resume versus a fresh worker within an unfinished task, with missing driver/fixture custody named as readiness inputs.

The 2026-10-15 20:48Z convergence correction is recorded in protocol.md (source record sha256 c75fd83da2664e94ff3064ff9d938ac36b9bde85a733803c3f0ebe4b6d98455c). Lead-note fields, size budgets, update conventions and relaunch heuristics are explicitly CANDIDATES, not requirements selected by CC design. The withdrawn 20:45Z definition is not a source of adoption authority. The paired experiment holds note conventions equal and tests history continuity only; it cannot qualify their exact format or budgets.

• Base commit: 8053f14e7e0817d2502dff2bb284bfb7e1b0152f
• Lane: lane:foundation
• Owned paths: evidence/artifacts/runtime-worker-context-engineering-20261015/ and its six registrations in manifests/evidence.json.

Research only. No installs, inference, Muse invocation, configuration/default/fence changes, catalog/ledger/research-state changes or comparison execution. Command-center context and orchestration rows remain out of scope.

SOTA sources

• earendil-works/pi 04566631c2b51fb688486bc1f040f60d1518c126 (1.1.0), packages/coding-agent/docs/{compaction,sessions,settings,configuration,rpc-commands}.md: native thresholds, structured summaries, session identity and accounting. Same-version installed documentation drift is explicitly retained.
• anomalyco/opencode 53d1eabb61e21162157817bf677da0a4ad3332e3 (qualified 1.18.35), packages/opencode/src/session/{overflow,compaction,llm}.ts, core summary template and vendor CLI/rules docs. Shared installed 2.0.27 is recorded as drift, not silently substituted.
• Meta's Muse Code documentation (interactive/configuration/extending/session-messaging), read 2026-10-15; pinned fence decisions 2026-10-14-muse-file-fence.md and 2026-10-15-muse-toolchain-directory-grants.md. Dated public docs do not prove behavior at native 1.4.7-R6101.1; numeric thresholds and telemetry completeness remain gaps.
• anthropics/claude-code 03b5892ee77de289bcd5dfa39f541525ade74aff:CHANGELOG.md, relevant 2.1.296–2.1.301 changes; vendor documentation for context-window, settings, memory, hooks, subagents, CLI and monitoring, read 2026-10-15. The installed 2.1.301 version is distinct from the practice skill's 2.1.295 record.
• Pinned primary sources for all accessible carried survivors, ECC, ai-memory and superpowers are linked file-by-file in survivors.md; Agent Console's inaccessible retained pin is a named gap. No vendor benchmark claim is treated as a measured worker saving.
• Existing layer-resolution protocol and wave/sweep inputs at the base; Add isolated Pi and OpenCode trial profiles and native invocation receipts #1033's trial profiles; New WSL clean install plan (NativeHorizonNY) and the host Claude config reconciliation #1174 Decisions 5–6 at open head 8b28bbda7dca372baa8026ca0bc6dce21b5b5626; existing native-practice planning_2b public result and hash-verified private custody references, named without host paths.

Evidence-class table

Local commands run

git fetch origin
git switch -c codex/runtime-worker-context-engineering-20261015 origin/main
python3 scripts/validate.py
  passed: 71 components, 12145 hashed_files, 4 profiles, 240 receipts
python3 scripts/host_name_scan.py evidence/artifacts/runtime-worker-context-engineering-20261015
  exit 0
git diff --check
  exit 0

Source/relative-link checks returned 33 verified pinned files, no unresolved targets and no host-home literals. Six files are registered through the existing host_receipts.register_file API. No new Python script or source-policy grant is needed. The final validation was run after the last source correction and registration.

Normal push also passed the repository's pre-push host-name scan, three registry coverage checks and published-digest privacy check.

Decision record

evidence/artifacts/runtime-worker-context-engineering-20261015/policy-trial.md is a candidate design for trial, not an adoption decision. comparison.md fixes the rule for the single unrun experiment, without making its controls production requirements. protocol.md records the convergence correction, corrects the supplied workers-angle description to its actual two INSUFFICIENT verdicts and distinguishes current native binaries from qualified profile pins.

The existing paired-run driver and sealed F5/F6 fixture manifest are missing custody inputs owned by cc-native-practice. The co-op confirmed their absence in the prior run directory. This PR supplies no invented execution command or replacement fixture.

Host evidence

No evidence/hosts/ receipt or platform status change.

Checklist

• No GitHub Actions or workflow changes.
• Public records contain no credentials, account fields, environment values or private host paths; directing text is not quoted.
• No new paid service, install or required secret.
• Peer-owned files/worktrees preserved; a clean new branch starts at the requested base.
• Muse fence and named-directory rules hold; no invocation or startup extension enabled.
• Catalog, ledger, research-state and runtime settings untouched; all policy labels are TRIAL.