Add strict parsing

Fixes sillsdev/serval#2

…), not nondet traversal

Task O2 (rust-optimizations-phase2.md): measurement-only, no engine fix. Profiled the two
known-slow Amharic words post-P10 (which doesn't touch Amharic's >64-def id-lane) with a real
profiler search first: WPR needs elevated privileges unavailable here; samply installs but only
records on Linux/macOS. Fell back to targeted Instant::now() phase instrumentation in the
HC_STEP_STATS style.

Added three permanent, opt-in (HC_FST_PROFILE=1), near-zero-cost diagnostics:
- hc_fst::profile (traverse.rs): Transduce::run/nondet-branch/distinct() call counts, total/max
  time, and traversed-set / raw-result-list sizes.
- hc_rules::morph::dedup_profile (morph.rs): push_remove_duplicates (the "keep-longer" dedup,
  lead #2) call counts, time, and out-list length.
- hc-cli prints both as FSTPROF/DEDUPPROF stderr lines alongside the existing STEPS line.

Findings (docs/o2-profile-findings.md): ~90-96% of wall-clock is inside Transduce::run for both
words, but the dominant sub-cost is NOT the nondeterministic traversal (14-30%) -- it's the
trailing distinct() dedup (59-81%), a Vec-scan + pairwise FstResult::result_eq comparison
processing up to ~500K raw results in a single call (C#'s Enumerable.Distinct is hash-backed
O(n); Register already derives Hash+Eq). Lead #2 (push_remove_duplicates) is refuted for these
words: <0.1% of wall-clock, out-list length never exceeded 1. Lead #3 (template-battery interior
memo) is inconclusive by direct measurement (no time went through that layer for either word) but
a code-reading check found hc-memo's template_memo already at parity with C#'s TemplateMemo
granularity. Live C# oracle run on ሌባዬ: 64.0s vs Rust's ~145s (~2.2-2.3x), byte-identical
signature -- confirms O2's "~2x per-step" framing freshly on this exact word.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…hash-backed Distinct) dominates the FST-traversal cost gap, not the nondeterministic traversal itself

Verified independently: build clean; clippy had 19 new missing_const_for_thread_local
warnings from the new profiling instrumentation, fixed mechanically via cargo clippy --fix
before merge (now 0 warnings, matching repo baseline); full suite green.

Findings (rust/docs/o2-profile-findings.md): on both known-pathological Amharic words, 89-96%
of wall-clock is inside Transduce::run, but within that the dominant cost is the trailing
distinct() post-processing dedup (59-81% of total wall-clock) -- an O(n x kept) Vec-scan +
pairwise result_eq over raw match lists reaching 327K-501K elements -- not the nondeterministic
traversal loop (14-30%). Register already derives Hash+Eq and looks hashable in practice, so a
hash-based dedup matching C#'s Enumerable.Distinct looks directly applicable. Lead #2
(push_remove_duplicates keep-longer dedup) refuted as negligible. Lead #3 (template-battery
memo) inconclusive by measurement but code-reading confirms parity with C#. Flagged for a
Fable interpretation/fix follow-up, not implemented here (profiling-only per task scope).

Confirmed lead #1 (Transduce::run), re-localized to distinct()'s O(n x kept) dedup rather
than the nondeterministic traversal itself; refuted lead #2; deprioritized lead #3. Next step
is a Fable-tier hash-based distinct() redesign, not yet started.