cli: choice parameter values reject hyphen/case aliases (sigmoid-novelty → sigmoid_novelty) #888

Phenomenon

Algorithm parameter names already fold case and -/_ (batch-size → batch_size) and a near-miss typo ends with did you mean 'batch_size'? (#876). Algorithm names (#874), --kind (#882) and --agent (#884) do the same.

Parameter values that are declared choices= do not. A user who writes the hyphenated or shouty spelling of a documented choice — --param sampling=sigmoid-novelty, --param archive_mode=keep-all, --param mode=COSINE, or the same strings in a spec's algorithm.params — gets a hard reject that only dumps the list. There is no fold and no did you mean …?, even for a close typo such as novlty.

This blocks a quick algorithms run smoke: the help text shows sigmoid_novelty / keep_all, and people naturally type the hyphenated form they use everywhere else.

Evidence

On current main (ebf39689b49e1ef43e8442dd86c2ee62b25307de):

• resolve_params(archive_selection, {"sampling": "sigmoid-novelty"}) → algorithm.params.sampling: must be one of ['sigmoid_novelty', 'novelty', 'performance', 'uniform', 'best'], got 'sigmoid-novelty'
• {"sampling": "SIGMOID_NOVELTY"} and {"archive_mode": "keep-all"} / {"archive_mode": "KEEP_ALL"} (dgm) and {"mode": "COSINE"} (skillopt) fail the same way
• {"sampling": "novlty"} dumps the list with no did you mean 'novelty'?
• Contrast: {"batch-size": 2} already resolves to batch_size (algorithms: unknown parameter names don't fold hyphens or suggest a near-miss (batch-size / bacth_size → batch_size) #876); algorithms show archive-selection already folds the card name (algorithms show/run: algorithm names are case- and hyphen-sensitive #696)

Shipped choice parameters today: archive_selection.sampling, dgm.archive_mode, skillopt.mode.

Expected

A string choice value that matches exactly one declared choice once case and -/_ are ignored resolves to that choice (sigmoid-novelty → sigmoid_novelty, KEEP_ALL → keep_all, Linear → linear). algorithms run … --param … and a spec's algorithm.params both store the canonical spelling (same Param.coerce path, so MCP / plugin hosts inherit it).

A remaining unknown that is a close typo of one choice names that guess (novlty → did you mean 'novelty'?). A nonsense value (greedy) still lists the choices and does not guess.

Acceptance

• resolve_params(archive_selection, {"sampling": "sigmoid-novelty"}) (and SIGMOID_NOVELTY) returns sampling == "sigmoid_novelty"
• dgm archive_mode=keep-all / KEEP_ALL → keep_all; skillopt mode=Linear → linear
• algorithms run archive_selection <prompt> <cases> --param sampling=sigmoid-novelty writes that canonical value in the spec
• sampling=novlty errors with did you mean 'novelty'?
• sampling=greedy still errors with must be one of [… tit] and no guess
• changelog.d fragment

Non-goals

• Folding spaces or other punctuation (sigmoid novelty, sigmoid.novelty)
• Changing what the algorithms do with a value once it is canonical
• Adding new choice parameters

Risk hint

T1/T2 (Param.coerce validation only; no evolve-loop change).

Related

#876 / #877 (parameter-name fold), #874 (algorithm-name fold / did-you-mean), #882, #884.