fix(validator): make decision-event matrix harness-aware

[chozandrias76]

Summary

Closes #113.

cupcake validate flagged every decision verb on a Cursor-native event (beforeShellExecution, afterFileEdit, beforeReadFile, …) as incompatible with an empty Supported: list, because the decision-event compatibility matrix only contained Claude Code events. Swapping the verb did not help — the matrix had no entry for the event at all, so the generated Cursor builtins failed validation out of the box.

Changes

• decision_event_matrix.rs: add Cursor's actionable native events, with verb sets derived from the Cursor response handlers (harness::response::cursor):
  • beforeShellExecution / beforeMCPExecution: Halt, Deny, Block, Ask (full allow/deny/ask permission gate)
  • beforeReadFile: Halt, Deny, Block (ask is coerced to deny)
  • beforeSubmitPrompt: Halt, Deny, Block (ask coerced; context injection unsupported)
  • stop: Halt, Block
  • Cursor's fire-and-forget after* events (always return {}) are intentionally omitted.
• Add DecisionEventMatrix::knows_event() to distinguish an unknown event from a known event with an empty verb set (e.g. SessionEnd).
• validator/rules.rs: skip events the matrix has no spec for, so the Claude-centric matrix no longer flags other harnesses' (or fire-and-forget) events, while known-but-empty events stay strict.

Tests

Added regression tests: Cursor before-events support the expected verbs; after* events are unknown (skipped); and knows_event() distinguishes unknown from known-but-empty. cargo test -p cupcake-core --lib validator → 27 passed.

Notes

• Scoped to the validator only; no engine/runtime behavior changes.
• Verified end-to-end: after this change, cupcake validate on a global config scaffolded by cupcake init goes from 5 errors to 0.

[chozandrias76]

For posterity, I reported this bug to my LLM (Claude) initially, I tasked it with investigation, and I gave it some criteria to validate for assumptions and to ensure the bug was real (according to my configurations). The text report (issue), PR body, and changes were LLM generated and validated by me (Colin)

[chozandrias76]

For posterity, I reported this bug to my LLM (Claude) initially, I tasked it with investigation, and I gave it some criteria to validate for assumptions and to ensure the bug was real (according to my configurations). The text report (issue), PR body, and changes were LLM generated and validated by me (Colin)

Oh and I also did a build test and smoke test with claude.