Limit wavediff filters to selected value streams

[jeet-dekivadia]

Fixes #20.

Summary

wavediff --filter already narrows the signal hierarchy before name/meta/value comparison, but the value-reader side still streamed filtered-out handles and dropped them later. This makes narrow filters correct in output, but less useful on large traces because the reader/channel/buffer path still sees data the user excluded.

This change derives per-reader local include sets from the filtered merged SignalMap and uses them while reading values:

• FST readers now pass the selected local handles through FstFilter.include
• VCD readers skip non-selected local handles before batching them into the diff channel
• multi-file --set1/--set2 offsets are preserved by translating merged handles back to each reader's local handle space

Fixture metrics

Using the existing fixtures as a deterministic scope check:

• tests/data/counter.fst: --filter '*.clk' narrows 4 signal names to 1
• {set_clk.vcd, set_counter.vcd}: --filter '*.clk' narrows 4 signal names to 1
• {set_clk.vcd, set_counter.vcd}: --filter '*.cyc_plus_one' narrows 4 signal names to 1, and the selected handle lives in the second reader, covering non-zero set offsets

After this patch, those narrowed scopes are reflected in the value stream instead of only in the final comparison/output layer.

Validation

cargo build --release
cargo test
cargo clippy -- -D warnings

I also checked the fixture scope counts with:

target/release/wavecat --names tests/data/counter.fst
target/release/wavecat --names --filter '*.clk' tests/data/counter.fst
target/release/wavecat --names tests/data/set_clk.vcd tests/data/set_counter.vcd
target/release/wavecat --names --filter '*.clk' tests/data/set_clk.vcd tests/data/set_counter.vcd
target/release/wavecat --names --filter '*.cyc_plus_one' tests/data/set_clk.vcd tests/data/set_counter.vcd

[toddstrader]

Thanks. Confirmed that this is way faster for small filter patterns. However, I'm a bit worried about reader_include_sets() in the default case. Does that assemble all the offsets when no -f is given? If so, can we short circuit it when there is no filter?

[jeet-dekivadia]

Good catch — the previous version did build those sets even without -f. I pushed 75c7e74 to short-circuit that path.

reader_include_sets() now compares the retained merged-handle count with the readers' native signal count and returns None when the hierarchy is unfiltered. That propagates as FstFilter.include = None; the VCD path also skips the per-change HashSet lookup entirely. Filtered and match-nothing cases still use explicit per-reader sets.

I added regressions covering the short circuit for single-file FST, single-file VCD, and multi-file VCD inputs. cargo test and cargo clippy --all-targets --all-features -- -D warnings pass locally, and both upstream CI jobs are green on the new commit.

[toddstrader]

Great, this seems marginally faster in the default scenario now. -f top.* is slower than no filter, but that would be self-inflicted.