Per-instance heavy-phase outputs cached for reuse across many
(tau, core_budget_fraction) selection cells. The selection /
post-pass / render pipeline then runs against this state cheaply.
--mode locate (#126): same heavy phase and the SAME selection as pack
mode, rendered as a ranked navigation list with provenance reasons and no
source bodies.
Heavy phase: clone/parse/fragment/discover/tokenize/score. Independent
of tau/core_budget_fraction. Designed to be computed ONCE per
instance and reused across an arbitrary number of selection cells.
The unified diff of diff_range as git prints it, minus the file sections
diff mode never discloses: secret-like paths, ignored paths, and lock files
(#112). Additive output for --with-raw-diff (#150) — it feeds no
selection state, so selection is bit-identical with and without it.
Selection + the 3 post-passes, shared verbatim by the pack renderer
(select_with_params) and the locate renderer — extracting it is pure
code motion so both modes select identically by construction.
Light phase: selection + 3 post-passes + render. Cheap. Re-runnable
against the same ScoredState with different (tau, core_budget_fraction)
to sweep a calibration grid without re-doing the heavy phase.