Deep-record heal 80% threshold blocks ~71 partially-degraded DeepPartial token ghosts — lowering it regresses #177 overload suppression

[jagguji]

Summary

The #205 fix recovers 350 → 111 of blend 0.0.38-beta.1's record→string token-config
ghosts. The residual ~71 (beta.0's floor is ~40) are partially-degraded records — a mix
of good fields and a few DeepPartial-deepened opaque token-config fields — and they are held back
by the deep-record heal's 80%-degraded threshold (healGhostRecords, src/extract.mjs):

if (!bad || bad < e.fields.length * 0.8) continue   // only heal a MOSTLY-fallback record

Instrumented distribution on blend 0.0.38-beta.1 (recordEntries=3194):

belowThresh(<80%)=406   ← the partial records holding the residual ghosts
attempted: healed=72, noImprove=182 (genuinely unmappable, correctly rejected), unbounded=0

Lowering the threshold to heal any record with ≥1 fallback field recovers all of them
(record ghosts 111 → 0, total loose 843 → 168), the bounded gate still rejects nothing wrongly,
and everything compiles.

Why it's not a one-liner — the threshold is load-bearing

Removing (or lowering) the threshold regresses the #177 overload-ambiguity suppression. Golden
generic-instantiation-distinct asserts that OrderDep.ambiguous (an overloaded slot) keeps a
flagged string for its _format field — suppressed via ctx.noPolyTag — while
unambiguous() gets the exact [#"ordCt"]. Its own comment states the contract:
ambiguous -> flagged, unambiguous -> [#"ordCt"]. With the threshold lowered, the heal re-resolves
the suppressed record at depth 0 without the per-slot suppression context, fakes an exact
[#"ordCt"] polytag in the ambiguous position, and merges the two records — a "flag-don't-fake"
violation.

A targeted attempt to skip records with a suppressed field (f.type.relinkNp) did not restore
the suppression — the suppression marker isn't carried where the heal can see it at the top-level
field, so the interaction needs to be untangled properly.

Fix direction

Recover the partially-degraded token ghosts without re-resolving deliberately-suppressed
(noPolyTag) fields. Options to evaluate:

1. Preserve suppressed fields across a heal — heal a record below the 80% threshold, but when
   accepting the rebuild, keep any field that was noPolyTag-suppressed in the original at its
   original (flagged) type instead of the rebuilt polytag. Requires the heal's captured _heal.ctx
   (or the field IR) to reliably carry the suppression so the rebuild re-suppresses or is
   field-wise overridden.
2. Field-level heal — instead of an all-or-nothing record rebuild gated on a global degradation
   ratio, re-resolve only the individual opaque (depth-truncated, non-suppressed) fields, leaving
   suppressed/other fields untouched. Sidesteps the threshold entirely.

Either way, add a golden case that pins BOTH: a partially-degraded DeepPartial token record whose
opaque field recovers, AND an ambiguous-overload record whose suppressed field stays flagged, so the
two mechanisms can't regress each other.

Impact

Cosmetic/fidelity only — the residual are flagged ⚪ loose string (usable, honest), not broken.
Closing this takes blend 0.0.38-beta.1 from 111 to ~40 residual (the genuinely-unmappable floor).
Follow-up to the #205/#206 fix.