perf: stabilize SymbolSlab inlining + reduce encode allocations

[virtuallynathan]

Why

1. Codegen fragility: The SymbolSlab accessor methods (get, get_mut, get_pair_mut) use #[inline] (soft hint). Under lto = "fat" + codegen-units = 1, LLVM outlines them when the impl block grows. Adding a single dead-code method to SymbolSlab causes a ~24% roundtrip regression from binary layout shifts alone.
2. Redundant copies in encode: SourceBlockEncoder stores source symbols as individual Symbol objects in a Vec<Symbol>, each owning a separate Vec<u8>. Building a SourceBlockEncoder copies every source byte twice: once into the Vec<Symbol>, then again into the intermediate SymbolSlab.

How

Commit 1: stabilize SymbolSlab accessor inlining under LTO

• Promote get(), get_mut(), and get_pair_mut() from #[inline] to #[inline(always)].
• These are the PI solver's hottest accessors. physical_index() was already #[inline(always)].
• 1 file, 3 lines changed.

Commit 2: reduce encode source symbol allocations

• Add SymbolSlab::from_bytes() to construct a slab directly from a contiguous byte slice with padding.
• Change create_symbols() to return a SymbolSlab instead of Vec<Symbol>, eliminating the per-symbol heap allocation.
• In create_d(), bulk-copy source bytes into the intermediate slab via copy_block_from() instead of iterating per-symbol.
• 2 files, +53/-37 lines.
• Public API unchanged.

Benchmarks (Zen4, EPYC 9654P, symbol_size=1280)

Back-to-back runs, cooldowns between switches. Criterion uses 100 samples per metric.

codec_benchmark (criterion, reliable)

Comparison is this branch (both commits) vs the inline-only baseline, measured back-to-back:

Metric	inline-only baseline	this branch	Delta
encode 10KB	15.86 us*	13.07 us	-17.6%
roundtrip 10KB	14.42 us	11.49 us	-20.3%
roundtrip repair 10KB	48.41 us	45.61 us	-5.8%

* Criterion's cached value used for comparison (reported as -17.6% change). Inline-only and master encode times were comparable (~13 us).

End-to-end vs master (separate back-to-back run):

Metric	master	this branch	Estimated total delta
encode 10KB	13.46 us	13.07 us	-3%
roundtrip 10KB	14.64 us	11.49 us	-21%
roundtrip repair 10KB	48.93 us	45.61 us	-7%

decode_benchmark (5% overhead, single-run harness)

K	master	this branch	Delta
10	3,122	3,084	-1.2%
100	3,906	3,906	0.0%
250	4,192	4,353	+3.8%
500	4,883	5,207	+6.6%
1,000	4,859	5,374	+10.6%
2,000	4,680	4,859	+3.8%
5,000	4,156	4,173	+0.4%
10,000	3,081	3,475	+12.8%
20,000	2,128	1,885	-11.4%
50,000	1,270	1,405	+10.6%

Commit 2 only touches encoder code. Decode variation is from the #[inline(always)] in commit 1 and single-run noise.

decode_benchmark (0% overhead, single-run harness)

K	master	this branch	Delta
10	3,303	2,959	-10.4%
100	4,110	4,029	-2.0%
250	3,693	4,353	+17.9%
500	4,066	4,217	+3.7%
1,000	3,876	4,197	+8.3%
2,000	3,983	4,250	+6.7%
5,000	3,299	3,742	+13.4%
10,000	2,933	3,277	+11.7%
20,000	2,128	1,961	-7.8%
50,000	1,372	1,449	+5.6%

Single-run harness. The 0% overhead path does not exercise the PI solver, so variation is noise.

Tests

• cargo clippy --all --all-targets -- -Dwarnings — clean
• cargo test --all — 60 passed, 4 ignored
• cargo build --features benchmarking,serde_support — clean
• cargo test --features benchmarking — clean

Commit history

1. 985e073 perf: stabilize SymbolSlab accessor inlining under LTO (src/symbol_slab.rs)
2. 54a5920 perf: reduce encode source symbol allocations (src/encoder.rs, src/symbol_slab.rs)

Notes

• Commit 1 is prerequisite for commit 2. Without #[inline(always)] on the accessors, adding from_bytes() to SymbolSlab triggers the codegen fragility described above, causing a ~24% roundtrip regression from layout shifts.
• The decode-side copy reduction from the original version of this PR was dropped — it showed marginal improvement indistinguishable from noise and added complexity.

[virtuallynathan]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Hooray!

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[cberner]

lgtm, but I had a couple minor comments

[cberner]

Should this be an assert!() rather than debug_assert!()?

[cberner]

It seems like this should be assert!(). It'd be a pretty serious bug if the mapping is Some, ya?

[virtuallynathan]

Moved asserts to runtime, not just debug builds.