fix(cli/scoring): make TextNormalizer replacement order deterministic (#921)

## Summary

Fixes #911 — the CI ASR benchmark scored byte-identical
hypothesis/reference pairs differently between runs (the
`1688-142285-0002.flac` record moved between WER 0% and 10% with no code
change).

## Root cause (proven, and stronger than suspected)

The issue guessed per-**process** dictionary hashing. It's
per-**instance**: Swift seeds each hash table from its storage
allocation, so two identical dictionary literals in the *same process*
iterate in different orders. Since `normalize()` builds fresh
`contractions`/`abbreviations` dictionaries on every call, the reference
and hypothesis calls inside one `calculateWERAndCER()` get independent
orders. With overlapping keys, `don't` expands to `do not` when `n't` is
applied first but `don not` when `'t` wins — so ref and hyp can disagree
within a single scoring call.

Evidence (standalone harness compiled from the unmodified
`TextNormalizer.swift`):
- 300 fresh processes on the exact #911 pair: **144 diverged** (`do not`
vs `don not`, exactly the 1-word/1-char gap in the artifact: WER 1/10,
CER 1/35).
- Two identical dict literals built back-to-back in one process:
different key order 5/5 runs.

## Fix

Apply all four replacement tables (`britishToAmerican`, `abbreviations`,
`contractions`, `numberWords`) sorted **longest key first**, lexical
tie-break. Deterministic, and the most specific rule now always beats
its substring — which is also the semantically correct expansion:

| input | before (order-dependent) | after |
|---|---|---|
| `don't` | `do not` **or** `don not` | `do not` |
| `can't` | `can not` **or** `ca not` | `can not` |
| `I'd been` | `i had been` **or** `i would been` | `i had been` |
| `it's been` | `it has been` **or** `it is been` | `it has been` |

Verified with the patched file across 50 fresh processes: byte-identical
output every run.

## Tests

`AsrTextNormalizerTests`: the exact #911 record scores WER 0/CER 0
across 100 repeated calls; repeated-call byte-determinism;
overlapping-key expansion expectations.

Note for reviewers: CI benchmark baselines may shift slightly on merge —
historical numbers were sampled from random orderings; the deterministic
ordering picks the correct expansions, so WER should move down or stay
flat.

Author: Alex-Wengg

Sources/FluidAudio/ASR/Parakeet/Unified/benchmark.md

@@ -5,6 +5,9 @@ Measured through the FluidAudio Swift managers (`UnifiedAsrManager` for batch,
 scored with the repo's `TextNormalizer` (same normalization as `asr-benchmark`).
 Encoder precision: **int8**. Run with `swift run -c release fluidaudiocli unified-benchmark`.
 
+> WER rows below predate the deterministic normalizer ordering (#911); reruns
+> may differ by a few tenths of a point with no model change.
+
 | Mode | Avg WER | Aggregate WER | Median WER | Median RTFx | Overall RTFx | Long files (>15s) |
 |------|---------|---------------|------------|-------------|--------------|-------------------|
 | batch | 2.16% | 1.68% | 0.00% | 130.0x | 143.6x | 238 |