fix(troubleshoot): shorten test fixture paths for Windows MAX_PATH

[MarinRzv]

Problem

Installing the UiPath plugin on Windows fails during checkout: Claude Code
git-clones the whole repo into ~/.claude/plugins/, and Windows rejects any
file whose absolute path exceeds MAX_PATH (260) without long-paths enabled.
The uipath-troubleshoot test fixtures were the offenders - worst repo-relative
path was 203 chars (175 paths >180), so with a typical ~80-90 char install
prefix they blew past 260 and broke the install.

Note: excluding tests/ from the npm "files" list does not help - the
marketplace clone brings the whole repo and fails at checkout, before any
packaging filter applies. The only fix is shorter paths in the repo.

Changes

Three commits, each independently reviewable:

1. Content-hash the fixtures + collapse the mock dir names.
   fixtures/mocks/responses/.json -> data/m/r/<sha1[:10]>.json,
   and the shared shim _shared/mock_template/mocks/ -> .../m/. Manifest
   "file" refs rewritten in lockstep; "match" keeps the real commands/GUIDs.
   Identical-content fixtures dedupe to one hash. Worst path 203 -> 138.

2. Fix a pre-existing dead manifest rule in cross-system/rpa-preflight-failure:
   two rules shared match "or folders list --output json"; the first
   (first-match wins) pointed at a fixture that never existed, erroring at
   runtime and shadowing the working rule. Removed the dead rule - the survivor
   resolves to the current-user folder listing (contains the Personal workspace).

3. Flatten the activity-packages/ tree. Drop the per-package subfolder
   (word-activities/, database-activities/, ...); all 164 activity scenarios
   now sit flat under activity-packages/ with the package encoded as a short
   name prefix (db-, cv-, excel-, gsuite-, mail-, o365-, py-, sys-, uia-, web-,
   word-, classic-). Fixes the _shared template path (../../../ -> ../../) in
   the 164 activity task.yaml files. Worst path 138 -> 121.

Also updated: the scenario generator (generate_scenario.py - emits data/m/r/
plus hashed names, --group activity-packages + token-prefixed scenario name),
coverage_report.py, verify_manifest_commands.py, and the
tests/tasks/uipath-troubleshoot/CLAUDE.md authoring docs.

Result

Worst repo-relative path 203 -> 121. With a ~100-char Windows install prefix
that lands at ~221, comfortably under 260.

Verification

• Reconstructed a sandbox and ran the shim end-to-end - resolves fixtures,
  exit 0, .calls.jsonl written.
• All 184 manifests / 1773 "file" refs resolve; 0 hash collisions; 0 leftover
  old-structure dirs.

Note

The test tree no longer mirrors the skill's
references/activity-packages/-activities/ playbook folders. This is
organizational only - nothing functional references the test scenario paths
(decision-tree resolution_ref/context point at the skill references/ tree,
which is untouched).

[github-actions]

PR Review: fix(troubleshoot): shorten test fixture paths for Windows MAX_PATH

Review in progress...

• Read review criteria and project conventions
• Analyze full diff (origin/main...HEAD)
• Check existing PR conversation
• Review modified files in detail
• Verify manifest integrity and link resolution
• Post review findings

View job run

[github-actions]

Claude finished @MarinRzv's task in 3m 20s —— View job

---

Coder-eval task lint (advisory)

184 task YAMLs changed (20 modified, 164 renamed); 0 Critical, 0 High, 0 Medium, 0 Low, 184 OK.

Rubric: .claude/commands/lint-task.md. This check is advisory and never blocks merge.

Evidence of passing run

✅ Author confirms passing run: "Reconstructed a sandbox and ran the shim end-to-end — resolves fixtures, exit 0, .calls.jsonl written. All 184 manifests / 1,773 file: refs resolve; 0 hash collisions; 0 leftover old-structure dirs."

Per-task lint

All 184 changed task YAMLs are verdict: OK. The changes are purely mechanical path adjustments — no test logic was altered:

• 164 activity-packages tasks (renamed): Moved from per-package subdirectories (e.g. word-activities/word-replace-text-file-locked/) to flat layout (word-replace-text-file-locked/). _shared path updated from ../../../ → ../../. Fixture dirs renamed fixtures/ → data/, mocks/ → m/.
• 20 non-activity-packages tasks (modified): Same fixture/mock directory renames (fixtures/ → data/, mocks/ → m/) applied in-place.

All tasks follow the mandated troubleshoot pattern:

• ✅ skill_triggered + llm_judge only (no forbidden criteria types)
• ✅ run_limits properly separated from agent: block
• ✅ No @uipath/cli in env_packages
• ✅ All tagged e2e (not smoke)
• ✅ Minimal prompts — no over-specification
• ✅ Canonical llm_judge block with include_reference: true + include_agent_output: true

Sampled 9 tasks across all groups (activity-packages, products/orchestrator, products/integration-service, products/maestro, runtime-exceptions, cross-system) — all structurally identical and conformant.

Within-PR duplicates

No duplicate clusters detected. The 164 renamed activity-packages tasks share a YAML scaffold but each exercises a materially distinct failure scenario (different activity, different error, different root cause). This is template reuse, not duplication.

Conclusion

✅ All 184 changed tasks pass the rubric. Evidence of passing run confirmed. Changes are mechanical path shortening with no test logic modifications.

---