lifter batch: score registered owners inside retained modules (--owner-spans, --sweep)

[wowwheaties]

Summary

lifter batch scores each retained module at its extent, but many modules hold more than one registered owner (regions.json manual_regions). Scoring the unchanged draft at the owner's own span instead often links exact already — that observation produced most of the owners in #6, by hand. This makes lifter do it.

• --owner-spans (batch, list): after the module pass, score every registered owner whose start lies inside the module at that owner's span; one row per owner (resource_3a2:02000924 span=412 differing_halfwords=6 (module 02000924 extent 1036)) and an owner_spans=… owner_exact=… summary line.
• --sweep: when the owner span is not exact, retry with the span widened to the rest of the module and report both (sweep_span=… best=owner|sweep), for owners registered short of the pool island their function really runs to (the 3c5:02001238 / 379:02000074 shape from ☀️ 74% – adopt seventy-four exact owners on 7d8a72083: eight main-image functions and sixty-six scene sequences #6).
• trailing pad: a reference exactly one halfword longer than the candidate with wrong_instructions=0 is ranked exact and marked trailing_pad=1 with the raw count kept in the row; adopt still decides on the real bytes.

Default output is byte-identical without the flags. No new dependencies; cargo test and cargo fmt --check clean; make verify green.

On this tree (9dbf29f)

lifter batch --all --jobs 8 --owner-spans --sweep: 187 modules, 273 owner spans, 47 exact from the unchanged draft (one of them via the pad rule: 3a6:020011a0), 14 spans where the widened sweep scores better than the registered span.

Notes

• Under --jobs 8 a handful of owner scores come back with no score line (23 here); each scores fine when retried alone. That is the existing scorer subprocess path under load, not new code — left as is rather than adding a retry you might prefer to shape.
• The sweep widens to the module end only. Widening to the next registered owner start would be a natural refinement if you want the tool to propose the corrected span for the registry.

[wowwheaties]

Independent check of the 47 differing_halfwords=0 owner rows from batch --all --owner-spans --sweep on 9dbf29f: each module draft re-scored one at a time with overlay score <draft> --owner <owner> --span <span> — 46/47 reproduce dh=0 as printed. The 47th is the trailing_pad=1 row (resource_3a6:020011a0 span=184): raw dh=1 at 184, dh=0 at 182, --align header class=layout-only wrong_instructions=0, i.e. the pad rule fires only where the registered span includes the word after the final return.

[wowwheaties]

Superseded by #8: lifter was removed in 6f712af, so this is re-done against thumb2c (same three pieces, same opt-in flags; 51 owner spans exact on 187bb97). Closing.