Surgical performance pass on hot paths

[edkerk]

Tier 3 of the raven-python review: a surgical performance pass — verified hot-path
footguns only, no redesign and no behaviour change.

Changes

• manipulation/add.py + manipulation/change.py — add_reactions_from_equations
  and change_reaction_equations resolved each equation token by re-scanning
  model.metabolites for a name / name[comp] match, i.e. O(R*k*M) for R reactions
  x k tokens on an M-metabolite model. Both now resolve through a shared
  (name, compartment) -> metabolite index (_build_met_index), dropping the bulk
  path to O(R*k). The index is seeded from the current mets (first match wins,
  exactly mirroring the old linear scan) and updated as new mets are created, so the
  existing cross-token / cross-reaction dedup is preserved.
• reconstruction/homology/homology.py — the ortholog filter used
  DataFrame.apply(lambda r: ..., axis=1), which builds a Series per row. Replaced
  with a comprehension over the two columns (the model_genes values are already
  sets, so membership is O(1)); matches the zip(..., strict=True) idiom already used
  a few lines below.
• analysis/sampling.py — the random objective was assembled with sum(terms),
  which re-canonicalises the optlang expression on every term (O(n^2) — the footgun
  documented and avoided everywhere else in the package). Switched to
  optlang.symbolics.add(terms), matching init/ftinit.py et al.

Not changed (verified no-op)

The plan's fourth perf item (trim model.copy() in the gap-fill / task loops) is
already done or load-bearing: init/taskfill.py:_feasible was converted to
with model: (see its comment — that copy had dominated gap-fill runtime), and the
remaining copies in gapfilling/fill.py build or return new models. Nothing safe left
to trim.

Tests / verification

Adds a cross-reaction metabolite-dedup regression test (the invariant the index must
preserve). pytest (619 passed), ruff check, and mypy (ad-hoc on the changed
sources) all green; no behaviour change on the existing suite.