Fix false-positive archetype classification and verification run count

[cursor]

Summary

Fixes three bugs in workflow_profile_service.py:

1. "doc" substring falsely matches "docker" (high severity)

_DOC_TEXT_HINTS contained "doc" which was matched via bare in substring check against lowercased text. This caused any session mentioning "docker", "docstring", etc. to be misclassified as "docs" archetype.

Fix: Replaced the hint tuple + substring matching with a compiled regex using word boundaries (\b). \bdocs?\b matches "doc" and "docs" as standalone words but not "docker" or "docstring".

2. "port" substring falsely matches "import" (medium severity)

_MIGRATION_TEXT_HINTS contained "port" which substring-matched "import", "export", "support", "report", etc. Combined with the weak file-count guard, many non-migration sessions would be incorrectly classified.

Fix: Same approach — compiled regex with \bport\b matches "port" as a standalone word but not when it appears inside other words like "import" or "export". Other migration stems use \w* suffix to preserve prefix-matching behavior (e.g., migrat\w* matches "migrate", "migration", "migrating").

3. Verification run count tallies types not actual runs (medium severity)

verification_run_count iterated over execution_types (a set of distinct type strings) instead of the full execution_evidence list. A session with 5 test runs would report count of 1 because the set collapses all "test" entries.

Fix: Changed iteration to loop over execution_evidence and extract each record's evidence_type, counting all matching records rather than just unique types.