Performance: cheaper Declaration ordering and less evidence-collection allocation
Two correctness-preserving performance changes, measured against a corpus of
real `.hie` files (the nix-ci3 project: 15 packages, 138 `.hie` files, ~8 MB),
built with the matching GHC. Output is byte-identical (same 459 weeds) before
and after, and the test suite passes.
## Changes
**1. Collect evidence uses directly instead of via `Data.Tree`**
`requestEvidence` built a `Tree [Name]` and immediately flattened it away with
`concat . Tree.flatten`. The tree was pure allocation overhead; collect the
names with a direct recursive concatenation. ~3.8% less total allocation, time
neutral.
**2. Compare declarations by `Unique` with a structural tie-break**
`Ord Declaration` was a *lexical* comparison of the module and occurrence names,
paid by every `Set`/`Map Declaration` operation (a hot spot in profiles).
Compare by the occurrence name's `FastString` `Unique` (an `Int`, and what its
`Eq` already uses) first, falling back to the original structural comparison
only to break `Unique` ties -- keeping the instance consistent with the derived
`Eq`. Since `Set Declaration` is now ordered by run-dependent `Unique`s, the
output sort gets a `displayDeclaration` tie-break so the reported order stays
deterministic.
Note: comparing the *module* by `Unique` as well was tried and was ~2.3x
**slower** -- a `Module`'s `Unique` is not a cached field, so `getUnique`
recomputes it on every comparison. Only the occurrence name's `FastString`
`Unique` (which is cached) is used.
## Measurements
| metric | baseline | after both | delta |
|--------|----------|-----------|-------|
| serial mutator time | -- | -- | **-15% to -24%** |
| total allocation | 14.48 GB | 13.72 GB | **-5.3%** |
| weeds reported | 459 | 459 | identical |
The mutator-CPU win is large and clean. The user-visible *parallel* wall-clock
gain is smaller and noisy on a many-core machine, where parallel GC dominates
the variance and masks the mutator improvement.
## Context
The previously dominant cost -- `getEvidenceTree` recomputed per (decl, name) --
is already handled by the `analyseEvidenceUses` cache on `master`. These two
changes pick up the next pieces. The evidence phase is still the single largest
chunk of the run and is largely untouched here.