nccgs 2.1.0 → 2.1.1

package/CHANGELOG.md

@@ -1,200 +1,235 @@
-# Changelog
-
-## 2.1.0 — Unreleased
-
-- Handle the final LOW diagnostic finding: empty, incomplete, unreadable or
-  disappearing admission locks refuse with an actionable message, preserving
-  the original error cause. Keep ownership checks and explicit retry unchanged.
-
-- Resolve the follow-up review's two low-severity observations: distinguish
-  admission contention from active uninstall, and record session recovery as an
-  immutable pending receipt plus a separate success/failure outcome. Missing or
-  malformed outcomes remain unconfirmed; prior attempts and reviews stay intact.
-
-- Resolve the first independent maintenance review's four findings: add explicit
-  hash-bound operator recovery for crashed launcher leases with immutable receipts;
-  retain unknown process-state refusal and all task/job/run ownership checks.
-  Serialize run/install/agent/migration/policy admissions against uninstall,
-  cancel pre-mutation refusals without stranded receipts, and return backup IDs
-  even when empty-directory cleanup produces warnings. Historical negative review
-  and packages stay unchanged; the corrected archive requires fresh qualification.
-
-- Add `nccgs uninstall --project DIR`, a read-only `--dry-run` and guarded
-  `--rollback ID`. Remove unchanged inventoried framework files and exact NCCGS
-  hooks/imports; preserve customized files, shared Claude preferences, game source
-  and the full `.nccgs` policy/task/review/evidence/budget history.
-- Save every removed/replaced byte in a project-local uninstall receipt before
-  changing files. Refuse unsafe paths, ambiguous custom hooks, active/unverified
-  work and rollback over later user edits; attempt restoration on write failure.
-- Register `nccgs run` sessions so uninstallation cannot race a new managed
-  launch. An interrupted uninstall blocks launch/update until recovery. No process
-  is killed and no task, acceptance requirement, review or budget is reset.
-- The published 2.0.0 archive remains unchanged. This feature receives local
-  framework and installation tests; it is not a new Claude Code/Unity benchmark.
-  See [uninstall and recovery](docs/UNINSTALL.md).
-
-## 2.0.0 — 2026-10-28
-
-Published to npm on 2026-10-28. The following notes describe its original
-preparation scope; the exact publication and native qualification have separate
-frozen receipts. No new version is published by preparing 2.1.0.
-
-- Publish existing external operator Editor start, graceful shutdown command and
-  settled stop receipts into immutable reviewer-readable project bundles. Preserve
-  original bytes, hashes and ownership, reject inconsistent/uncertain/forced stops,
-  and identify historical imports without rewriting sealed tasks or verdicts.
-- Keep RC4 handoff, scene guard, source controls, process management, independent
-  review, acceptance and total-budget enforcement unchanged. No new game gate or
-  process permission is introduced.
-- Record scoped exact-RC4 native qualification: real dirty-scene refusal, new repair,
-  automatic PARTIAL handoff with no continuation, all 28 criteria/385 assertions,
-  independent PASS and normal closure. The original LOW evidence-location finding
-  remains in that sealed review; its addendum is assessed separately.
-- Stable archive validation and independent maintenance review are recorded in
-  release receipts. They do not claim another full Unity task through these bytes.
-
-
-## 2.0.0-rc.4 — 2026-10-28
-
-Prepared candidate; unpublished. No npm tag or stable promotion.
-
-- Settle a protected implementation work phase with a deterministic PARTIAL
-  report when current implementation and measured checks are ready for review.
-  A prose-only response no longer spends another continuation solely on handoff.
-  Preserve independent verdicts, all acceptance gates, history and original caps.
-- Pin an atomic loaded-scene precondition to new Unity repair reservations.
-  Dirty named, additive or untitled scenes block the operation without auto-save
-  or discard. Retain failed attempts and old reservations; no implicit retry.
-- Reserve final independent review money/time within the original root budget.
-  Share the work pool across design, slices, repair and verification. Admit
-  unknown worker cost only through verified, exclusive native session coverage;
-  missing session usage fails closed, and resumed cumulative cost is never summed
-  twice. The prior native reserve case qualified its disclosed new-script and
-  asset repair path with 28 criteria/385 assertions and independent final PASS.
-- RC4-specific handoff/scene changes receive framework tests and operator Unity
-  checks. A fresh native task through this exact archive remains pending.
-
-- Qualify the full component lifecycle on a fresh Unity project through actual
-  Claude Code: reviewed design, two dependent implementation slices from API
-  scaffolds, first-import verification, independent review, a disclosed fault,
-  typed repair and independent re-review. All 24 original criteria (349 assertions)
-  pass; the real failed PlayMode attempt remains recorded. The combined design and
-  implementation phases used 7/12 dispatches, three independent reviews and
-  USD 10.214066/12 within the original 75-minute envelope. Installed source guards,
-  process ownership, role limits, criteria and history remained unchanged. This
-  known-test component qualification does not establish broad game productivity.
-- Add opt-in exact `unityGeneratedFiles` declarations before task initialization
-  so Unity's first import can create IDE/local metadata without a false source
-  mutation failure. Pin the declaration as a canonical input and retain output
-  before/after hashes in managed receipts. Keep all undeclared files tracked;
-  reject source paths, folders, globs and canonical overlaps. Existing policies,
-  historical attempts, process ownership, review and budgets remain unchanged.
-- Settle cold Unity compilation without waiting for the qualified Roslyn server's
-  idle keepalive. After the command root exits and no other worker remains, request
-  graceful shutdown only from the verified compiler pipe inside the owned Windows
-  Job Object. Match the tested compiler/host binary hashes; unknown builds retain
-  the original wait/timeout behavior. Still drain every descendant and preserve the
-  command exit code, deadlines, independent review, criteria, history and budgets.
-- Avoid Unity CLI's two-minute idle auth broker in managed verification command
-  trees using the supported process-local `UNITY_NO_AUTH_BROKER=1` setting.
-  Preserve OS process ownership, descendant draining, failure/timeout semantics,
-  account settings, independent review, acceptance criteria and spent budgets.
-- Add bounded task slices with preview, pinned plans, dependency tracking and
-  focused worker prompts inside one original task. Final integration retains all
-  original criteria/checks and a reviewer independent of every contributor.
-- Protect aggregate dispatch capacity, original repair/QA/review reserves and
-  review limits. Preserve plan revisions, partial attempts and dependency history;
-  retries do not reset limits or create child/session budgets. Cost/time admission
-  stays unchanged.
-- Qualify the two-slice and existing-file `unity-code` repair path with actual
-  Claude Code workers and Unity tests on 2026-10-27. The controlled regression
-  retains all 24 criteria, failed evidence, both independent reviews and original
-  total budgets. This known-component case does not qualify large-game productivity,
-  new-file imports, Editor asset operations or the exact published RC archive.
-- Add opt-in schema-2 Unity repair packets: exact code edits/creation or pinned
-  Editor asset operations. Preserve schema-1 limits and historical reservations.
-- Bound Editor operations to one project, fixed positional arguments, finite
-  foreground calls and durable single-attempt receipts. Check source scope and
-  existing GUIDs; import new metadata before completion. Uncertain operations
-  retain ownership until externally reconciled, without refunds or automatic retry.
-- Unity repair still requires original measured checks, independent review and
-  manual/external acceptance. Synthetic tests do not qualify native Unity delivery.
-- Synchronize architecture and orchestration guidance with standalone defaults;
-  generate native role ceilings from agent-library frontmatter alongside routing.
-- Extend `task dispatch-prompt` to prepare registered implementation, analysis and
-  review assignments with safe PARTIAL handoff templates. Repair packets keep
-  their existing prompt and permissions.
-- Add read-only `task handoff-validate` for a worker's own declared handoff and
-  coordinator inspection, so schema errors can be corrected before completion.
-- Add coordinator `task record-review` to import a completed reviewer's exact
-  pinned result through the existing record gate. Review independence, objective
-  criteria, external/manual gates, task history and spent budgets are preserved.
-- The scoped native qualifications above predate the final RC4 handoff/scene
-  additions. Their frozen source, tests and historical evidence remain separate.
-
-## 2.0.0-rc.3 — 2026-10-07
-
-Published-1.3.0 upgrade correction on next; latest remains unchanged.
-
-- Fixed false overrides when a role generated by a 1.3.0 profile becomes a managed
-  file in 2.0. Existing generated fingerprints now identify unchanged agent files;
-  later user edits and explicit overrides still require review.
-- RC2 qualification exposed the issue in quality, balanced and inherit profiles.
-  The astra-claude path was unaffected. Original failed results are retained.
-- Release qualification uses the published 1.3.0 archive and the exact candidate
-  archive across four model profiles, owner settings and open task history,
-  override preservation, dry runs, explicit execution migration, rollback and
-  refusal of changed configuration/backups. See docs/UPGRADE-QUALIFICATION.md.
-- A new task executed by Claude Code through this exact release remains pending;
-  this correction does not promote stable or change the frozen game evaluation.
-
-## 2.0.0-rc.2 — 2026-10-07
-
-Release documentation follow-up on next; latest remains unchanged.
-
-- Unified current version labels and exact-version npm install, doctor, run and migration commands.
-- Replaced the obsolete Vietnamese Astra/Fable-default guide with standalone RC instructions.
-- Separated file update, explicit execution migration and guarded rollback; documented that rollback is not a full package downgrade.
-- Corrected active role counts and current workflow/project-policy guidance.
-- Added a release-status matrix and visible historical labels instead of hidden comments.
-- Retained RC1, frozen evaluation evidence and internal history. No runtime JavaScript, game source, model routing or acceptance threshold changed.
-- Packaging validation covers public contents, local documentation links, version consistency and fresh install/doctor. Published-1.3.0 upgrade qualification and a fresh exact-RC task remain pending.
-
-## 2.0.0-rc.1 — 2026-10-07
-
-Release channel: next. This candidate does not replace the stable/latest tag.
-
-### Candidate scope since npm 1.3.0
-
-- Standalone Claude Code execution with separate model, execution and external-review policies.
-- Reviewed design handoffs, pinned contracts, measured acceptance criteria and reviewer-owned verdicts.
-- Durable worker reservations, path ownership, protected repair/QA/review capacity,
-  verification jobs and bounded continuation without refunding spent attempts.
-- Explicit execution migration, preserved legacy obligations, backup and guarded rollback.
-- Reviewer capacity of up to 200 native turns; valid finding enums and standalone
-  command syntax documented to address observed closeout failures.
-- Portable public operating principles replace project-specific internal memory
-  in the package. The original development history is archived outside the package.
-
-### Validation scope
-
-The maintained framework regression recorded 183 PASS, 0 FAIL and 6 Windows
-symlink-fixture SKIP. An operator-assisted native Claude Code delivery V2 case
-completed 24 acceptance criteria,137 EditMode tests,52 PlayMode tests and an
-independent unconditional PASS. Failed attempts and evidence limitations were
-retained. This is not a clean autonomous benchmark or a stable-release claim.
-
-The game case ran a frozen framework with a session-only reviewer turn-capacity
-override. This RC packages the maintained source. Upgrade qualification from
-published 1.3.0 and a fresh end-to-end task using this exact RC are pending.
-Try a disposable project or backed-up copy; see UPGRADING.md.
-# Stable preparation follow-up — 2.0.0 (unpublished)
-
-- Correct remaining current-version RC labels while retaining historical release
-  records and their original qualification scope.
-- Add opt-in, correlated native-result transport for read-only analysis/review
-  handoffs when the native client requires returned text. Copies exact output;
-  existing review, acceptance, source, process and budget gates still apply.
-- Failed audit attempts and their consumed allowances remain in their original
-  history. New qualification and publication require their own explicit evidence.
-
+# Changelog
+
+## 2.1.1 — Prepared for release (2026-11-03)
+
+- Prepare the accepted local R11 r05 source as 2.1.1 because npm already contains
+  a different 2.1.0 artifact. Update package/version markers and current usage
+  guidance; runtime scripts, policy, model routes, gates and budgets are unchanged.
+  Qualification of this exact patch artifact is recorded separately. This entry
+  records preparation, not publication or an npm-tag change.
+
+### Included changes from the accepted local 2.1.0 qualification
+
+- Correct the stale hook-identity test by naming all 13 existing managed hook
+  identities, including the earlier Unity CLI-only `unity-transport-guard.mjs`
+  entry in `settings.fragment.json`. No hook or identity function is added or
+  changed by the standalone effort revision. Preserve exact legacy-command,
+  custom-hook and failure-policy assertions. The strict deep-route fixture uses
+  the existing deep-class Unity implementer; the gameplay programmer remains
+  standard-class in `studio.json`.
+
+- Synchronize the shipped synthetic benchmark route controls and disclosed
+  review fixtures with the standalone defaults. Exercise strict observed xhigh
+  acceptance, high/xhigh mismatches and unknown-effort refusal without inferring
+  provider execution or resetting spent dispatches.
+
+- Update the existing `claude-standalone` defaults: lightweight/FAST requests
+  Sonnet/high; standard requests Opus/high; deep/CONTROLLED and leadership request
+  Opus/xhigh; independent review alone requests Fable/high. The primary session
+  keeps its existing standard-class binding and therefore requests Opus/high.
+  Other profiles, gates, quotas and native role ceilings are unchanged. These
+  are configured requests, not observed model versions or effective effort;
+  missing native effort evidence remains UNKNOWN. The source stays unpublished.
+
+- Handle the final LOW diagnostic finding: empty, incomplete, unreadable or
+  disappearing admission locks refuse with an actionable message, preserving
+  the original error cause. Keep ownership checks and explicit retry unchanged.
+
+- Resolve the follow-up review's two low-severity observations: distinguish
+  admission contention from active uninstall, and record session recovery as an
+  immutable pending receipt plus a separate success/failure outcome. Missing or
+  malformed outcomes remain unconfirmed; prior attempts and reviews stay intact.
+
+- Resolve the first independent maintenance review's four findings: add explicit
+  hash-bound operator recovery for crashed launcher leases with immutable receipts;
+  retain unknown process-state refusal and all task/job/run ownership checks.
+  Serialize run/install/agent/migration/policy admissions against uninstall,
+  cancel pre-mutation refusals without stranded receipts, and return backup IDs
+  even when empty-directory cleanup produces warnings. Historical negative review
+  and packages stay unchanged; the corrected archive requires fresh qualification.
+
+- Add `nccgs uninstall --project DIR`, a read-only `--dry-run` and guarded
+  `--rollback ID`. Remove unchanged inventoried framework files and exact NCCGS
+  hooks/imports; preserve customized files, shared Claude preferences, game source
+  and the full `.nccgs` policy/task/review/evidence/budget history.
+- Save every removed/replaced byte in a project-local uninstall receipt before
+  changing files. Refuse unsafe paths, ambiguous custom hooks, active/unverified
+  work and rollback over later user edits; attempt restoration on write failure.
+- Register `nccgs run` sessions so uninstallation cannot race a new managed
+  launch. An interrupted uninstall blocks launch/update until recovery. No process
+  is killed and no task, acceptance requirement, review or budget is reset.
+- The published 2.0.0 archive remains unchanged. This feature receives local
+  framework and installation tests; it is not a new Claude Code/Unity benchmark.
+  See [uninstall and recovery](docs/UNINSTALL.md).
+
+## 2.0.0 — 2026-10-28
+
+Published to npm on 2026-10-28. The following notes describe its original
+preparation scope; the exact publication and native qualification have separate
+frozen receipts. No new version is published by preparing 2.1.0.
+
+- Publish existing external operator Editor start, graceful shutdown command and
+  settled stop receipts into immutable reviewer-readable project bundles. Preserve
+  original bytes, hashes and ownership, reject inconsistent/uncertain/forced stops,
+  and identify historical imports without rewriting sealed tasks or verdicts.
+- Keep RC4 handoff, scene guard, source controls, process management, independent
+  review, acceptance and total-budget enforcement unchanged. No new game gate or
+  process permission is introduced.
+- Record scoped exact-RC4 native qualification: real dirty-scene refusal, new repair,
+  automatic PARTIAL handoff with no continuation, all 28 criteria/385 assertions,
+  independent PASS and normal closure. The original LOW evidence-location finding
+  remains in that sealed review; its addendum is assessed separately.
+- Stable archive validation and independent maintenance review are recorded in
+  release receipts. They do not claim another full Unity task through these bytes.
+
+
+## 2.0.0-rc.4 — 2026-10-28
+
+Prepared candidate; unpublished. No npm tag or stable promotion.
+
+- Settle a protected implementation work phase with a deterministic PARTIAL
+  report when current implementation and measured checks are ready for review.
+  A prose-only response no longer spends another continuation solely on handoff.
+  Preserve independent verdicts, all acceptance gates, history and original caps.
+- Pin an atomic loaded-scene precondition to new Unity repair reservations.
+  Dirty named, additive or untitled scenes block the operation without auto-save
+  or discard. Retain failed attempts and old reservations; no implicit retry.
+- Reserve final independent review money/time within the original root budget.
+  Share the work pool across design, slices, repair and verification. Admit
+  unknown worker cost only through verified, exclusive native session coverage;
+  missing session usage fails closed, and resumed cumulative cost is never summed
+  twice. The prior native reserve case qualified its disclosed new-script and
+  asset repair path with 28 criteria/385 assertions and independent final PASS.
+- RC4-specific handoff/scene changes receive framework tests and operator Unity
+  checks. A fresh native task through this exact archive remains pending.
+
+- Qualify the full component lifecycle on a fresh Unity project through actual
+  Claude Code: reviewed design, two dependent implementation slices from API
+  scaffolds, first-import verification, independent review, a disclosed fault,
+  typed repair and independent re-review. All 24 original criteria (349 assertions)
+  pass; the real failed PlayMode attempt remains recorded. The combined design and
+  implementation phases used 7/12 dispatches, three independent reviews and
+  USD 10.214066/12 within the original 75-minute envelope. Installed source guards,
+  process ownership, role limits, criteria and history remained unchanged. This
+  known-test component qualification does not establish broad game productivity.
+- Add opt-in exact `unityGeneratedFiles` declarations before task initialization
+  so Unity's first import can create IDE/local metadata without a false source
+  mutation failure. Pin the declaration as a canonical input and retain output
+  before/after hashes in managed receipts. Keep all undeclared files tracked;
+  reject source paths, folders, globs and canonical overlaps. Existing policies,
+  historical attempts, process ownership, review and budgets remain unchanged.
+- Settle cold Unity compilation without waiting for the qualified Roslyn server's
+  idle keepalive. After the command root exits and no other worker remains, request
+  graceful shutdown only from the verified compiler pipe inside the owned Windows
+  Job Object. Match the tested compiler/host binary hashes; unknown builds retain
+  the original wait/timeout behavior. Still drain every descendant and preserve the
+  command exit code, deadlines, independent review, criteria, history and budgets.
+- Avoid Unity CLI's two-minute idle auth broker in managed verification command
+  trees using the supported process-local `UNITY_NO_AUTH_BROKER=1` setting.
+  Preserve OS process ownership, descendant draining, failure/timeout semantics,
+  account settings, independent review, acceptance criteria and spent budgets.
+- Add bounded task slices with preview, pinned plans, dependency tracking and
+  focused worker prompts inside one original task. Final integration retains all
+  original criteria/checks and a reviewer independent of every contributor.
+- Protect aggregate dispatch capacity, original repair/QA/review reserves and
+  review limits. Preserve plan revisions, partial attempts and dependency history;
+  retries do not reset limits or create child/session budgets. Cost/time admission
+  stays unchanged.
+- Qualify the two-slice and existing-file `unity-code` repair path with actual
+  Claude Code workers and Unity tests on 2026-10-27. The controlled regression
+  retains all 24 criteria, failed evidence, both independent reviews and original
+  total budgets. This known-component case does not qualify large-game productivity,
+  new-file imports, Editor asset operations or the exact published RC archive.
+- Add opt-in schema-2 Unity repair packets: exact code edits/creation or pinned
+  Editor asset operations. Preserve schema-1 limits and historical reservations.
+- Bound Editor operations to one project, fixed positional arguments, finite
+  foreground calls and durable single-attempt receipts. Check source scope and
+  existing GUIDs; import new metadata before completion. Uncertain operations
+  retain ownership until externally reconciled, without refunds or automatic retry.
+- Unity repair still requires original measured checks, independent review and
+  manual/external acceptance. Synthetic tests do not qualify native Unity delivery.
+- Synchronize architecture and orchestration guidance with standalone defaults;
+  generate native role ceilings from agent-library frontmatter alongside routing.
+- Extend `task dispatch-prompt` to prepare registered implementation, analysis and
+  review assignments with safe PARTIAL handoff templates. Repair packets keep
+  their existing prompt and permissions.
+- Add read-only `task handoff-validate` for a worker's own declared handoff and
+  coordinator inspection, so schema errors can be corrected before completion.
+- Add coordinator `task record-review` to import a completed reviewer's exact
+  pinned result through the existing record gate. Review independence, objective
+  criteria, external/manual gates, task history and spent budgets are preserved.
+- The scoped native qualifications above predate the final RC4 handoff/scene
+  additions. Their frozen source, tests and historical evidence remain separate.
+
+## 2.0.0-rc.3 — 2026-10-07
+
+Published-1.3.0 upgrade correction on next; latest remains unchanged.
+
+- Fixed false overrides when a role generated by a 1.3.0 profile becomes a managed
+  file in 2.0. Existing generated fingerprints now identify unchanged agent files;
+  later user edits and explicit overrides still require review.
+- RC2 qualification exposed the issue in quality, balanced and inherit profiles.
+  The astra-claude path was unaffected. Original failed results are retained.
+- Release qualification uses the published 1.3.0 archive and the exact candidate
+  archive across four model profiles, owner settings and open task history,
+  override preservation, dry runs, explicit execution migration, rollback and
+  refusal of changed configuration/backups. See docs/UPGRADE-QUALIFICATION.md.
+- A new task executed by Claude Code through this exact release remains pending;
+  this correction does not promote stable or change the frozen game evaluation.
+
+## 2.0.0-rc.2 — 2026-10-07
+
+Release documentation follow-up on next; latest remains unchanged.
+
+- Unified current version labels and exact-version npm install, doctor, run and migration commands.
+- Replaced the obsolete Vietnamese Astra/Fable-default guide with standalone RC instructions.
+- Separated file update, explicit execution migration and guarded rollback; documented that rollback is not a full package downgrade.
+- Corrected active role counts and current workflow/project-policy guidance.
+- Added a release-status matrix and visible historical labels instead of hidden comments.
+- Retained RC1, frozen evaluation evidence and internal history. No runtime JavaScript, game source, model routing or acceptance threshold changed.
+- Packaging validation covers public contents, local documentation links, version consistency and fresh install/doctor. Published-1.3.0 upgrade qualification and a fresh exact-RC task remain pending.
+
+## 2.0.0-rc.1 — 2026-10-07
+
+Release channel: next. This candidate does not replace the stable/latest tag.
+
+### Candidate scope since npm 1.3.0
+
+- Standalone Claude Code execution with separate model, execution and external-review policies.
+- Reviewed design handoffs, pinned contracts, measured acceptance criteria and reviewer-owned verdicts.
+- Durable worker reservations, path ownership, protected repair/QA/review capacity,
+  verification jobs and bounded continuation without refunding spent attempts.
+- Explicit execution migration, preserved legacy obligations, backup and guarded rollback.
+- Reviewer capacity of up to 200 native turns; valid finding enums and standalone
+  command syntax documented to address observed closeout failures.
+- Portable public operating principles replace project-specific internal memory
+  in the package. The original development history is archived outside the package.
+
+### Validation scope
+
+The maintained framework regression recorded 183 PASS, 0 FAIL and 6 Windows
+symlink-fixture SKIP. An operator-assisted native Claude Code delivery V2 case
+completed 24 acceptance criteria,137 EditMode tests,52 PlayMode tests and an
+independent unconditional PASS. Failed attempts and evidence limitations were
+retained. This is not a clean autonomous benchmark or a stable-release claim.
+
+The game case ran a frozen framework with a session-only reviewer turn-capacity
+override. This RC packages the maintained source. Upgrade qualification from
+published 1.3.0 and a fresh end-to-end task using this exact RC are pending.
+Try a disposable project or backed-up copy; see UPGRADING.md.
+# Stable preparation follow-up — 2.0.0 (unpublished)
+
+- Correct remaining current-version RC labels while retaining historical release
+  records and their original qualification scope.
+- Add opt-in, correlated native-result transport for read-only analysis/review
+  handoffs when the native client requires returned text. Copies exact output;
+  existing review, acceptance, source, process and budget gates still apply.
+- Failed audit attempts and their consumed allowances remain in their original
+  history. New qualification and publication require their own explicit evidence.
+
+## Unreleased R07 working alignment
+
+- Enforce direct Unity CLI/Pipeline transport at session launch and project hooks; default MCP discovery is isolated.
+- Reject Unity MCP tool calls and ordinary setup/launch commands, including for implementation workers and compatibility sessions.
+- Preserve explicit non-Unity integrations, native permission/model choices, upstream skill bytes and all historical budgets/verdicts.
+- Refuse unsafe legacy supervised recovery rather than silently rewriting retained arguments.

package/CLAUDE.md

@@ -1,8 +1,10 @@
-# NCCGS project bootstrap
-
-@MEMORY.md
-
-@.claude/nccgs/constitution.md
-
-Project-specific decisions live in `.nccgs/project.yaml`. Keep this file small;
-put local additions in that policy file or in canonical project documents.
+# NCCGS project bootstrap
+
+@MEMORY.md
+
+@.claude/nccgs/constitution.md
+
+Project-specific decisions live in `.nccgs/project.yaml`. Keep this file small;
+put local additions in that policy file or in canonical project documents.
+
+For Unity work, load `.claude/nccgs/protocols/unity-boundary.md` before any generic Unity Skill. Its CLI-only transport policy is mandatory.