[hzhou-muse]

Pull Request Description

A collective selection JSON, the built-in one or a user's MPIR_CVAR_COLL_SELECTION_JSON_FILE, could be wrong without anyone noticing. Algorithm parameters were keys such as "k=2", read with atoi, so a typo silently became 0 and the algorithm ran with it. Thresholds were read the same way. A wrong parameter count, an unknown algorithm, collective, condition or subtree name, or a malformed key hit an assert in MPI_Init. A tuning file that is subtly wrong is worse than one that is rejected.

This PR makes the selection JSON checked input: every value has a type, and every error is reported with its line and column.

Parameters are now written as "name": value, e.g. "algorithm=MPIR_Bcast_intra_tree": {"tree_type": "kary", "k": 4}. Each parameter's type comes from the CVAR in its cvar_params, so the JSON and the CVARs can't drift apart:

CVAR type	JSON value
int	an integer
boolean	true / false
enum	one of the CVAR's value names, as a string

The old "k=2" form is no longer accepted. coll_selection.json had one parameter, which is converted.

• Tree types: the *_TREE_TYPE CVARs for bcast, ibcast, allreduce, iallreduce and ireduce become enum CVARs, with values checked at compile time against MPIR_Tree_type_t. This fixes the topology tree types, which silently ran as kary, and removes a duplicate MPIR_CVAR_BCAST_TREE_TYPE declaration.
• JSON parser: it accepts integers, true and false, records each key's position, and provides typed getters.
• Generated code: gen_coll.py emits type-checked parameter parsing. For conditions it emits only a table; the parsing code lives in csel.c.
• Errors, not asserts: unknown names, missing, unknown or mistyped parameters, and malformed thresholds are all reported as errors. Thresholds must be integers that fit in an int.
• Structural checks: an algorithm= key must be the only key in its object (before, it leaked the conditions parsed before it and ignored the keys after it), a subtree object must not be empty (before, it asserted at run time), and a subtree name must fit in CSEL_MAX_NAME (before, it was copied without its terminating NUL). ! together with <, <=, > or >= (e.g. !comm_size>8) is rejected as ambiguous; !comm_size(8) still works.
• Every case covered: every object must end with an "any" key, and a new "error" leaf (e.g. "any": "error" after the built-in collective= list) ends a search on purpose. A search that loops through call= is stopped at run time. Before, a search that matched nothing asserted at run time, and a call= loop hung the collective. Now an "error" leaf or a loop makes the collective return "Collective selection failed to find an algorithm".
• Docs: Collective_Selection.md and cvar.md are updated.

Author Checklist

• Provide Description
• Commits Follow Good Practice
• Passes All Tests
• Contribution Agreement

🤖 Generated with Claude Code

[hzhou-muse]

test:mpich/ch3/tcp ✔️