Repair Windows staging and public skill discovery

[infotradescout]

Windows text staging translated line feeds while receipts expected exact UTF-8 bytes. This change writes with newline="", preserves read-only originals when a transaction is rejected, and cleans only owned staging files using supported Windows permission semantics. A narrow Git attribute also keeps the new digest-bound model capture byte-identical during Windows checkout.

The same candidate repairs public packaging and discovery through their existing owners. Direct activation and adjacent-consent behavior are preserved, as are the free portable core and one public SKILL plus seven role references. Worker, Objector and final-report instructions now carry applicable requirements, required states, concrete challenges, deferred work and proof limits through completion. One independently reviewed evaluation criterion removes an unrequested stone-designer demand from a seed that asked for stone browsing and functional cabinet/metal-building configuration. Earlier negative observations are retained unchanged.

Validation at frozen source 639ada169526fa0597397f8654ed21bf1d346922:

• Current role-mediated synthetic behavior: 34/34 captured responses, 216 criteria, independently graded with separate pre/post Objectors. The complete capture passes the unchanged evidence validator. Two fresh suite contexts share an Orchestrator; this is not 34 isolated cold starts, application execution, provider attestation or client parity. A separate local held-out assessment passed 12/12 responses across 94 criteria.
• The retained b9 full local gate passed five checks, including 316 unit tests and one host skip. All 41 canonical execution/test Git objects remain identical at this revision; that gate was reused rather than relabeled as a new full run. The changed instructions/projections passed 26 affected tests, native/discovery freshness checks and current package checks.
• Newly built portable/personal/public archives passed integrity, exact activation and content checks: 119/58/60 entries. Personal/public each contain one SKILL and seven role references, at 817,649/806,119 uncompressed bytes below the unchanged 1 MiB budget. Actual independent final review checked archive bytes against frozen sources. No archives have been uploaded as a production release.
• An isolated actual Git checkout-filter test reproduced LF-to-CRLF hash breakage and proved the exact LF attribute preserves the capture SHA256 without changing the validator or broad JSON handling.

The public doctor still reports three readiness errors: candidate-public proof status, genuine free-client conformance and public distribution status. The current model-evidence error is resolved. Historical free-client failure remains failed; no passing free-client, repeated supported-client cold-start acceptance, hosted CI, platform installation or live adoption is claimed. Existing hosted CI never started because GitHub reported an account billing lock; local Windows/Python 3.11 evidence does not establish hosted Ubuntu/Python 3.12. Production delivery is already owner-authorized, but these proof/access dependencies remain.

The published plugin remains separately identified as 1.0.5, and download wording keeps 1.0.8 archives as pending publication. This public candidate excludes the separate, compatible installed task-entry rollout at 6b0f54cf2dd350cdabb8f3bed3af2fb543ce7bd3; that local rollout is not public/free-client acceptance. The optional Infinity caller is separately reviewed/committed in an isolated SI branch and is excluded from this public candidate; neither a source pin nor a hosted dependency is introduced here.

The README-only follow-up at 3015ede746c84dfb5deac94a622b650701a21e15 makes CLI updates conditional on supported commands, applies checksum verification only to published archives, and documents the existing-plugin publisher update path. The canonical skill, adapters and plugin submission trees are unchanged from frozen 639ada1, so those archive proofs are reused without rebuilding or rerunning passed evaluations.