do not mutate caller-supplied metadata in Parameter and Dimension

[jmccreight]

Problem

Strict-mode Parameter and Dimension construction alias the
per-entry sub-dict of the caller-supplied metadata
(self.meta = meta[name]). Two code paths then write through that
alias into the caller's dict:

• Parameter.__init__: the creation-time bounded-maximum resolution
  writes meta['bounded_dimension_name'] and overwrites
  meta['maximum'] (dimension name -> numeric size).
• Dimension: the size setter writes meta['size'].

The first Parameters instance built from a metadata dict works and
poisons it; any later instance sharing that dict crashes in
Parameters.add() on bounded parameters, which reads
metadata[name]['maximum'] expecting a dimension name:

ValueError: `5` does not exist in metadata     # Dimension(name=5)

Repro (second iteration raises on current development):

import pyPRMS as pp
meta = pp.MetaData(verbose=False).metadata
for _ in range(2):
    p = pp.Parameters(metadata=meta, verbose=False)
    p.dimensions.add(name='one', size=1)
    p.dimensions.add(name='npoigages', size=5)
    p.dimensions.add(name='nobs', size=5)
    p.add(name='outlet_sta')
    print(meta['parameters']['outlet_sta']['maximum'])  # 'nobs' -> 5

Found integration-testing pywatershed against development ahead of the
next release (as discussed): pywatershed holds a module-level
MetaData().metadata and builds multiple Parameters from it; 2 of 6
DomainSubset tests failed this way.

Fix

Copy the metadata sub-dict at strict-mode construction in both
classes (self.meta = dict(meta[name])). Instances keep their
resolved values (bounded_dimension_name, numeric maximum,
size); the caller's dict is never modified. The non-strict paths
are untouched -- they take ad-hoc dicts and internal code deliberately
aliases meta between objects (e.g. ParamDimensions from global
dimensions).

Testing

• New regression test: two Parameters instances sharing one
  MetaData dict, adding a bounded parameter to each; asserts the
  instance sees the resolved bound and the shared dict is unchanged.
• Full tests/func suite: 303 passed.
• Downstream: pywatershed's test_domain_subset.py goes 4/6 -> 6/6
  against this branch (packaging 26.3 environment).

[pnorton-usgs]

Thanks for finding this — the root cause you identified is correct. I've pushed an alternative fix on branch fix/deepcopy-metadata-isolation that addresses it at a higher level in Parameters.__init__ rather than in each leaf class.

The approach uses two separate deepcopy calls:

1. __full_metadata — pristine copy, never mutated. Used by create_subset() to build child Parameters instances that need the original unresolved metadata (dimension names in bounded maximum fields, original dimension sizes, etc.).

2. _working_metadata — separate working copy that self.metadata, self.__dimensions, and all Parameter instances are free to mutate (bounded resolution, dimension size changes, etc.).

Since the caller's dict is never passed into the internal machinery, the shallow copies in Parameter.__init__ and Dimension.__init__ become unnecessary (though harmless if left in place).

This also prevents a subtler issue: create_subset() passes self.__full_metadata to a new Parameters(...). If __full_metadata shared sub-dicts with self.metadata (which gets mutated during bounded resolution), the child instance would receive already-resolved metadata and crash. The two-copy approach guarantees __full_metadata stays pristine.

All 302 tests pass. See: development...fix/deepcopy-metadata-isolation

[jmccreight]

Agreed. See my comment about keeping the test and then you can close the PR. Thanks!

[pnorton-usgs]

Issues addressed in this PR have been handled by PR #63 and #65. Closing this PR.