perf: resolve source spans through a line index

[e54-bot]

Summary

Program::validate resolved every recorded span by scanning the retained source text from the beginning for each line/column → byte-offset conversion (SourceDocument::byte_range → byte_offset). The identifier and nested-value spans added in #326 roughly tripled the number of checked spans, so this repeated full-text scan dominated large real-project workloads.

SourceDocument now records each line's start offset once; position resolution scans only the addressed line.

Measurements (bastion.ow, 447 709 bytes; 22 139 values / 3 051 actions / 303 rules)

stage	v1.0.0	v1.2.0	this branch
program.validate (release)	~3.00 s/op	~8.95 s/op	~10.3 ms/op
program.validate (debug)	—	~145.6 s/op	~119 ms/op
element_count (release)	~3.05 s/op	~8.91 s/op	~15.5 ms/op
element_count (debug)	—	—	~157 ms/op
parse / emit	~123 ms / ~10 ms	~134 ms / ~11 ms	~126 ms / ~11 ms

The per-node public span-path traversal that the issue hypothesized (condition_value_span / action_argument_value_span root-to-node lookups, 22 139 nodes) measures ~1.2 ms total in release — it was not the regression, and no parallel cache/index was added for it. The 1.0.0 baseline was already affected by the same scan; #326 amplified it into the minutes-scale dogfood regression.

Falsification / regression protection

• byte_offsets_match_full_document_scan: indexed resolution vs. the pre-index full-scan algorithm over a CRLF/Unicode document including boundary and overshot positions.
• byte_offsets_scan_only_the_addressed_line: bounds the bytes examined per resolution to one line — fails deterministically (no wall-clock threshold) if full-document rescans return.
• tests/perf_large_project.rs: ignored stage-level benchmark on the real bastion.ow input (same convention as audit_benchmarks), runnable via cargo test -p workshop-rs --release --lib perf_large_project -- --ignored --nocapture.

Verification

• cargo fmt --all --check ✅
• cargo clippy --workspace --all-targets -- -D warnings ✅
• cargo test --workspace --all-targets ✅ (lib 241, integration 107, cli 15+15, catalog-gen 9; 4 ignored benchmarks)
• cargo run -p workshop-rs --bin workshop-catalog-gen -- check ✅
• git diff --check ✅
• Existing provenance/span coverage passes unchanged (source_map, program_model, source_preservation, parser::spans_are_preserved).

Known remaining similar pattern (out of scope)

settings/schema.rs::source_value_range has the same per-call scan but serves single SettingSource::source_edit calls on a cold &str; it is not on the validate hot path and would need a different API to share an index.

Fixes #331

Generated with Devin