feat: mark Creed outputs as generated (COD-452)

[shivros]

Summary

• emit deterministic, target-owned .gitattributes blocks marking Creed-rendered outputs as linguist-generated
• preserve user rules, escaped literal paths, CRLF checkouts, disabled-target metadata, and per-file emission safety
• add dogfood attribution for Creed's own generated outputs and integration coverage

Validation

• go generate ./... && git diff --exit-code
• go build ./...
• go test -race -count=1 ./...
• go vet ./...
• test -z "$(gofmt -l .)"
• scripts/check-generated.sh

Closes COD-452.

[shivros]

🤖 Automated Review Panel

Reviewer A

Initial review identified stale/disabled metadata coverage, attribution after per-file output failure, missing dogfood output, then final review identified CRLF preservation and literal Git-attribute path escaping. All findings were addressed before this PR: generated blocks are CRLF-safe, preserve user content under uniform CRLF, escape literal paths, are only emitted after successful content emission, and dogfood output plus regression coverage are included.

Reviewer B

Independently identified the same CRLF-preservation and path-pattern correctness risks. The final implementation preserves uniform CRLF content, rejects mixed endings rather than rewriting user bytes, and escapes attribute patterns so whitespace/glob characters cannot broaden rules.

Validation

go generate ./... && git diff --exit-code; go build ./...; go test -race -count=1 ./...; go vet ./...; gofmt clean; scripts/check-generated.sh.

[shivros]

Auto-Merge Gate

Confidence: 0.91

The implementation directly satisfies COD-452: it deterministically emits target-owned .gitattributes blocks for Creed outputs while preserving user-authored rules, path literal semantics, CRLF handling, disabled-target behavior, and safe per-file emission ordering. The diff includes integration coverage and dogfood output; the prior dual-model review panel found and verified fixes for the relevant edge cases.

Checks observed: CI / Test, Lint, Creed context drift, and Secret Scan — all completed successfully.

Scope limits: eight files, confined to generated attribution support and its tests; no credentials, deployment, destructive behavior, or manual-review marker detected.