Release reviewed learning, local command pools and tool setup in forge 0.2.3

[ancongui]

Runs can lose current evidence eligibility or execution ownership between a cached-result check and its return. Keep the exact learning record locked through admission, validate the retained policy/source/outcome/expiry after the final actor check, and recheck current runner ownership before returning cached command results. Preserve the original WAL receipts and require an explicit refresh or reconciliation instead of another provider dispatch.

Add repository-scoped learning and explicit one-shot flycanon projection, plus administrator-configured local command pools with immutable profiles, native quotas, mutually authenticated controllers and bounded source/artifact transfer. The external-tool wizard separates connection saving, discovery, schema review and project grants; GitHub read tools are executable, while the other provider guides show unmet prerequisites. Refresh the illustrated documentation with 228 source-bound captures and prepare version 0.2.3.

SCM rotating-token uncertainty and confidential client/host retargeting now fail closed. MCP completion and cached-result paths recheck current authority. Native pyfly web/data/security/OAuth extras use the verified public 26.10.2 package. General-provider OAuth remains on a separate development branch.

Validation: final local native backend 1,180 passed with zero failures/errors/skips and unchanged source/services; Studio 71 unit and 168 browser cases plus lint/types/build; 41 docs and five repository-path browser cases; strict docs, link/discovery and privacy checks. Independent cache/transport review passed 18 cases. A fresh installed wheel passed CLI/assets and native authenticated API checks; all 185 runtime entries match source/wheel/installation, and the corrected source distribution matches 927 repository files. Actual SDK, CLI and stdio MCP agree on nine reads; the native handshake exposes 96 tools. Production source matches the independently reviewed immutable snapshot. The current tree adds four test-only files/changes to resolve the Ubuntu fixture failures described below. Earlier failed harness runs remain retained.

Local controller qualification covers a trusted Factory host and isolated command containers. Full coding-provider/browser pool isolation, Kubernetes, multi-host placement, live provider/SCM account acceptance and complete generated-product assembly remain separate work. Application deployments and deployment tests remain local.

The first remote run failed 20 runner cases because test fixtures hardcoded a macOS Docker path. Resolve the actual installed executable with the shared test helper, keeping production validation unchanged and adding no skips. The exact missing-path cases reproduced 20 failures before correction and passed all 20 afterward; 147 related cases and four prerequisite/production-negative controls passed locally. Current exact-head CI gates merge and release.