[Birfy]

Closes #876

What changed

Unknown algorithm parameter names now behave like algorithm names: a key that matches exactly one declared parameter once case and -/_ are ignored resolves to it (--param batch-size=1 writes batch_size), and a remaining close typo appends ; did you mean 'batch_size'?. Unrelated names like nope still get the full list and no guess. CLI --param, specs, and MCP share resolve_params.

Verification at the tip

Tip SHA: ea77629
Ran:

• pytest -q tests/test_algorithms.py tests/test_cli_algorithms_run.py tests/test_mcp_algorithm_spec.py tests/test_algorithm_entry_points.py tests/test_skill_library_example.py tests/test_skill_library_kind.py tests/test_changelog_fragments.py tests/test_demo.py → passed
  Not run:
• full suite — CI

[Birfy]

Review at ea776292e0a4b2da5be06cf8a2ead31087eada4c

Verdict: PASS

• Parameter aliases reuse the existing name folder (_fold / _resolve_alias); batch-size and NOISE_BAND resolve silently, and a collision of both spellings of the same parameter is refused.
• Suggestion only fires for a single unknown key that is a close match (cutoff 0.75) to the folded name; nope gets none, so existing regex/substring assertions stay valid.
• algorithm_override writes the canonical key into the spec, and the evoskill batch_size=1 starter default folds first so --param batch-size=2 is not treated as a duplicate.
• Tests: alias fold, collision, near-miss, no-guess, CLI hyphen --param, and CLI typo hint; algorithm / run / MCP / skill_library / demo / changelog suites passed locally.

risk_gate: pending (this comment).

[Birfy]

Verdict: PASS
Tip SHA: ea77629
Ran: pytest tests/test_algorithms.py tests/test_cli_algorithms_run.py tests/test_cli_algorithms_bare.py tests/test_mcp_algorithm_spec.py tests/test_algorithm_entry_points.py tests/test_cli_init_policy_slot.py tests/test_skill_library_kind.py tests/test_changelog_fragments.py at the tip → 215 passed; manual probes of resolve_params(evoskill, …) with Batch-Size (resolves to batch_size), bacthsize and batchsize (both end with "did you mean 'batch_size'?"); risk_gate.py from main → T2.
Not run: full suite (CI covers 3.9–3.13).

Correctness. _canonical_param_map reuses _resolve_alias, so parameter names fold exactly like algorithm names (#696): case and -/_ only, and an ambiguous fold stays unknown instead of silently picking one. Collisions after folding (batch_size + batch-size) are refused with a message naming both spellings, which is better than last-write-wins. The near-miss hint only fires for a single unknown key and uses the same 0.75 cutoff as #874, so nope still gets the plain list. All three entry points end up canonical: evolvespec goes through resolve_params, and the CLI's algorithm_override canonicalises before it builds the written spec, so --param batch-size=2 writes batch_size (the test checks the spec on disk).

Design fit. Good catch in starter_spec: without canonicalising first, setdefault("batch_size", 1) next to a user-supplied batch-size would have produced a collision error on init. A collision there is swallowed and then re-raised by algorithm_override, so the user still sees it.

Tests/docs. These cover the alias, the collision, three typo/meant pairs across evoskill, skillopt and rrsi, the no-guess case, and the CLI end to end, including "no spec written on error". The changelog fragment is present.

Non-blocking nits:

• cli.py now calls the private algorithms._canonical_param_map from two places. If more callers turn up, consider a public name, for example canonical_params, added to __all__.
• The import difflib inside the function is fine but differs from module-top imports elsewhere. Just a style point.