fix(skill): align author-saf-technique with the research validator

[bishnubista]

Summary

Aligns the author-saf-technique skill and the blank attestation template with what scripts/validate-technique-research.py actually enforces, and fixes one path and two portability problems in the clean-room instructions.

What changed

• Required prohibited_inputs fragments documented. The validator requires six exact substrings in prohibited_inputs for generation_mode: clean_room (validate-technique-research.py:519-530). Four of them were absent from the skill, and previous conversation appeared nowhere in the skill or template, so an attestation written in the skill's own wording failed validation. references/clean-room-generation.md now lists the six strings verbatim, and research/templates/technique/clean-room-attestation.yml carries them as a commented block to uncomment when a packet is converted to clean_room (the standard-mode scaffold keeps its not-applicable placeholder, since those inputs are permitted in standard reruns).
• Canonical paths corrected. SKILL.md listed traceability-ledger.yml as a bare path among otherwise-complete canonical paths; it lives at research/techniques/SAF-TXXXX/traceability-ledger.yml. references/methodology.md now names techniques/SAF-TXXXX/detection-rule.yml, the one packet artifact outside the research packet.
• GHSA search reconciled with clean-room mode. references/breach-and-vulnerability-research.md mandated a GitHub Security Advisories search that clean-room-generation.md forbids; the bullet now carries the clean-room exception (open an exact advisory URL only after non-GitHub discovery).
• Portable scratch root and a single retry convention. The frozen-bundle path was hardcoded to /private/tmp/..., which does not exist on the ubuntu-latest CI runner. It is now $SAF_CLEANROOM_ROOT, defaulting to ${TMPDIR:-/tmp}/saf-all-cleanroom, and must resolve to an absolute path outside the checkout. Attempt directories are created fail-closed (plain mkdir, next -rN on collision) so stale or concurrent attempts cannot mix under one FREEZE.sha256; freeze provenance is recorded in the bundle's integration-notes.yml. Retries and normalization re-freezes use SAF-TXXXX-rN; existing attestations had independently invented six different suffixes (-normalized, -retry, -r2, -retry3, -validator, -isolated).
• "Strict validator" wording. No --strict flag exists; strict checking is the default and --draft disables it. Wording now says "default (non---draft) mode." The --draft help string said it only skipped "completion statuses"; it now states which strict-only checks it relaxes (including accepting the placeholder zero history SHA) and that schema, cross-file joins, and core clean-room assertions remain enforced.

Testing

• python3 -m unittest discover -s scripts -p "test_*.py" — 15 tests, OK
• validate-technique-research.py --all, validate-framework-model.py, generate-technique-catalog.py --check, validate-detection-registry.py, generate-detection-coverage.py --check, generate-mitigation-catalog.py --check — all exit 0
• Reproduction of the original defect: in a scratch copy, rewriting SAF-T1309's prohibited_inputs in the skill's verbatim wording fails with four prohibited_inputs missing ... errors; the unmodified packet passes.
• Template verified both ways in a scratch copy: a standard scaffold from scripts/new-technique.py produces zero clean-room errors; the same scaffold converted to clean_room per the documented procedure produces zero prohibited_inputs missing errors.