fix(backfill): merge-preserving regen, real recipe yields, single-flight resolve

[icemansparks]

Fixes three defects found in a code review of the price-backfill toolchain.

Silent override loss on in-place regeneration

backfill.py --out-sql (which defaults to --existing-sql, documented as in-place regeneration) emitted only rows derived from the current candidates run. Any currently-overridden item absent from candidates.csv (BoP gear, legendaries, or simply a narrower candidate query) silently disappeared from the regenerated SQL, reverting to default AH pricing. write_outputs() now seeds the output from the existing overrides and overlays only items present in the new records, mirroring derive.py's merge. Regression test added.

Recipe yield hardcoded to 1

spelldbc.py recorded every crafting recipe with yield 1, so multi-output recipes (e.g. basePoints=4 -> 5 per cast) were over-priced by the yield factor in recipe_prices.py, and the error propagated recursively into higher-tier recipes using the output as a reagent. Yield is now extracted as EffectBasePoints + 1 (the 3.3.5a minimum-yield formula; variable rolls use the minimum, so derived prices are never inflated). Non-positive yields (bogus data) skip the recipe with a stderr warning instead of silently pricing at yield 1. Tests cover yield-1, fixed multi-yield, recursive propagation, and the negative-basePoints skip.

Lock held across network I/O

BuildIdState.note_result held the state lock across the sentinel fetch and buildId re-resolve, collapsing worker concurrency to 1 during every redeploy check. The lock now guards only shared state; a _resolving flag (held until the buildId update completes) keeps the re-resolve single-flight. Concurrency tests added (verified to fail against the pre-fix code).

Test suite: 51 passed.