Extract scheme-selection policy into a pluggable CostModel

[connortsui20]

Tracking Issue: #8701

Rationale for this change

The compressor's selection objective — maximize estimated compression ratio — is hardcoded into the selection loop, so it cannot be varied per workload (#7697). This PR makes the objective a pluggable CostModel: schemes continue to produce model-independent estimate facts, the model converts those facts into Cost, and the compressor selects the lowest-cost candidate.

The default SizeCost preserves today's ratio ordering and canonical-acceptance threshold. It prices the exact ratio signal (Cost(-ratio), canonical baseline Cost(-1.0)) rather than recomputing bytes, so IEEE rounding cannot turn a strict ratio ordering into a cost tie.

What changes are included in this PR?

Details

• Candidate: an opaque, borrowed view of the mechanical facts available for each selection option, including its estimated ratio and optional compressed sample. Sampling exposes the compressed sample array to the model only while that candidate is priced; the selector does not retain it or clone cascade history.
• vortex_compressor::cost: a finite, totally ordered Cost, the CostModel trait, and the default SizeCost. Cost is the only candidate-ranking value; compression ratios remain estimator input and tracing metadata.
• EstimateScore is removed. Deferred estimate callbacks now receive best_ratio: Option<f64> directly, and sampled estimates represent a zero-byte output as an absent ratio. The Delta and Sequence schemes are updated to the simplified callback contract.
• The selection loop is wired to the model. Installing a custom model disables compression-ratio early-exit hints for deferred callbacks, since a custom ordering may prefer a lower-ratio candidate.
• The initial model boundary and the decisions that remain outside it — constant handling, AlwaysUse, the final byte-acceptance gate, and extension-versus-storage selection — are documented explicitly.
• The winner_compress trace span gains model-defined estimated_cost alongside estimated_ratio.
• Selector-level tests prove that a custom model can change the winner, candidates must beat canonical_cost, sampled arrays are exposed only for sampled candidates, and custom models are not incorrectly pruned by ratio thresholds. The golden encoding-tree and size snapshots remain unchanged.

This PR does not add model-driven pruning, scheme priors, Delta-specific policy, builder/session plumbing, or a production time-based model.

What APIs are changed? Are there any user-facing changes?

This adds cost::{Cost, CostModel, SizeCost, Candidate} and CascadingCompressor::with_cost_model.

It deliberately changes the scheme-facing deferred callback API: EstimateFn receives Option<f64> (best_ratio) instead of Option<EstimateScore>, and EstimateScore is removed. Scheme::expected_compression_ratio itself retains the same signature. The default SizeCost preserves existing selection ordering, and the golden encoding-tree and array-size snapshots remain unchanged.

[codspeed-hq]

Merging this PR will not alter performance

⚡ 3 improved benchmarks
❌ 1 regressed benchmark
✅ 1626 untouched benchmarks
⏩ 42 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	compact_sliced[(4096, 90)]	779.7 ns	867.2 ns	-10.09%
⚡	Simulation	bitwise_not_vortex_buffer_mut[128]	273.6 ns	244.4 ns	+11.93%
⚡	Simulation	compact[(4096, 90)]	867.2 ns	779.7 ns	+11.22%
⚡	Simulation	compare_int_constant	298.4 µs	268.4 µs	+11.15%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing cost-model-r2-costmodel (0228bc7) with develop (42bea5d)}

Footnotes

1. 42 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩