Invert the framework updater from a denylist ("framework owns everything unless
preserved") to an explicit allow-list manifest ("operator owns everything unless
framework"). A path the manifest never anticipated resolves to operator-owned by
the fail-safe default, so it is structurally unreachable by any write or prune.
Root cause (#791): `mosaic update` re-seeds via `install.sh` keep-mode, whose
`rsync -a --delete` + hand-maintained PRESERVE_PATHS denylist wiped operator
paths the denylist forgot (agents/*.conf, policy/*.md, *.local.md, harvester
SOP, tools/_lib/credentials.json, unanticipated fleet files).
- framework-manifest.txt: single SSOT ([framework]/[operator], deny-wins,
UNKNOWN=>operator fail-safe), read by BOTH installers.
- src/framework/manifest.ts: pure resolver (parse/matchGlob/resolveOwnership/
frameworkSubtreeRoots/planPrune) — the testable seam.
- tools/_lib/manifest.sh: bash resolver (compiled globs, fork-free hot path),
sourced by install.sh; parity-tested against the TS resolver.
- install.sh keep mode is now manifest-driven (no --delete): overlay-copy
framework files, scoped-prune only retired framework files inside shipped
subtrees. Operator + unknown paths are never written or deleted.
- file-ops.syncDirectory gains an isOperatorOwned guard; file-adapter derives it
from the shared manifest, replacing the drifted hardcoded preservePaths.
Tests (TDD, red->green):
- HARD GATE test-upgrade-manifest-guard.sh: 10 operator sentinels (incl. an
unanticipated one) survive a keep-mode reseed byte-identical + mtime-unchanged;
retired framework file pruned; secret value absent from output. RED 31 fail on
the old installer -> GREEN 48 pass. Wired merge-blocking into CI.
- manifest-parity.spec.ts (§6.1): bash<->TS agree on 34 paths + subtree roots.
- manifest.spec.ts: 18 tests incl. planPrune property test + shipped-tree
completeness (§6.2).
- test-install-migration.sh F6 flipped: an unanticipated operator fleet file now
MUST survive keep-mode reseed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Each step was re-running pnpm install independently, and all quality
steps (typecheck, lint, format, test) ran in parallel. On merge commits
with more accumulated code this pushed the CI runner over its memory
limit (exit code 254 = OOM kill).
Fix:
- install once, share node_modules via Woodpecker workspace volume
- sequential execution: install → typecheck → lint → format → test → build
- corepack enable in each step (fresh container) but no redundant install
