fix: classify apply_patch failures

[sorted-ai-bot]

Summary

• Makes apply_patch adapter non-interactive with patch -f so malformed or mismatched patches return diagnostics instead of hanging or leaking raw patch prompts.
• Classifies common failures: plain non-git diff headers, missing target files, interactive strip-level prompts, stale hunks, already-applied patches, malformed patches, missing patch binary, and timeouts.
• Adds integration coverage for success, plain headers, file-not-found, and stale hunk cases.

Tests

• npm run build in adapters/
• ./gradlew test --no-daemon in java/
• go test -count=1 -run TestApplyPatch_Classifications -v ./integration
• go test ./cmd/mcphub

Scope

• Bugfix-only public Helm patch.
• Does not sync MCPHUB internal alpha features such as capability_gap, result cache, dry-run, or task-context filtering.

[sorted-ai-bot]

@YumaKakuya PR #15 merge request

[sorted-ai-bot]

@YumaKakuya conflict resolved after PR #14 merge. Rebased/merged main into bot/helm-fix-apply-patch-diagnostics, kept both TestSessionOpenRecoveryTool and TestApplyPatch_Classifications, and re-ran local tests: npm run build, ./gradlew test --no-daemon, ./gradlew jar --no-daemon, go test ./cmd/mcphub, go test -count=1 -run 'Test(SessionOpenRecoveryTool|ApplyPatch_Classifications)' -v ./integration.