Keep explicit signal filters when expanding a search

[00200200]

Fixes #75.

expand() built each derived signal leg with replace(query, text="", **override). When the caller had already set --symbol (or --file / --error / --test), that overwrite dropped their constraint. The JSON filters object still echoed the original flag, so the page claimed a scope the query no longer applied.

Appending the derived value is not a fix either: values inside one signal filter are ORed, which would widen the call rather than keep it.

This skips a derived leg of a kind the caller already filtered on. The text leg and the filters-only leg still carry that evidence; cross-kind filters (e.g. --file on a symbol leg) are unchanged.

Tested with pytest tests/integration/test_expansion.py tests/unit tests/integration/test_search.py.

[tarekziade]

Thanks that looks right

One subtlety: rank_spec() gathers structured signals from the surviving legs, so skipping a leg also removes its derived signal from scoring. The term remains in the lexical query, but loses its signal-overlap weight.

Could we retain derived signals for ranking independently of which legs execute?

The regression test should also assert constrained_numbers == {2} It already indexes a valid matching thread, but the current assertions would pass if the constrained search returned nothing.

[00200200]

Thanks, both points were right. Pushed a follow-up:

• Ranking: rank_spec() now also folds in the derived terms for any kind the caller already scoped, so skipping that leg no longer drops the term's signal-overlap weight. It only widens the scoring question; the filters (and therefore the scope) are unchanged. Kinds the caller left open still arrive through their own leg, so nothing is counted twice. Added test_a_skipped_derived_leg_still_counts_toward_ranking (it fails without the change) plus a no-double-counting case.
• Regression test: the constrained call now asserts constrained_numbers == {2}, so it can no longer pass by returning nothing.

Full suite passes locally (864 passed) and ruff is clean; the Postgres half of the parity tests is skipped here since I don't have a Postgres URL configured.