feat(linkage): ADR-0032 Stage A — cell-first calls for baked-record names (KIND-8 reloc)Verified

Widen the cave#266 cell-first call emission from pkgsym names to the
name-keyed world: a warm global call to a name with a baked canonical
record now loads the record base (imm64, new KIND-8 baked reloc:
consumer re-patches to ITS record base or an idempotent mint; the
+32 fn-cell displacement lives only in the code — codex P1-1), tests
the fn-cell, calls through it, and falls back to the name-keyed rel32
call on an unbound/nil cell.  Stage A is a pure emission widening:
every fallback patch is still recorded and the re-walk healing is
untouched; redefinition now ALSO heals instantly through the cell
(the ADR-0031 install door writes exactly this record).

Gates, each fixing a measured witness/producer divergence:
- runtime unification only (cold/self-host emission byte-identical);
- fully-booted sessions only (*ag-use-hosted-macro-eval*): ADR-0026
  turns unification on BEFORE prelude bring-up, and prelude emission
  feeds the .fco witness offsets, which must stay producer-identical;
- never during bundle emission (*ag-fco-bundle-emit*, set by BOTH
  ADR-0017 witness producers — ag-cold-bundle-build natively,
  produce-warm-bundle.lisp via the compiled setter): eligibility
  consults the producer's own static pool and funcl-cold's is a
  subset of stage2's, so per-name agreement is impossible;
- the baked lookup routes through a runtime HOOK
  (*ag-call-baked-cell-fn*, the *ag-reader-find-package-fn* pattern)
  set from stdlib, keeping the emit path free of static edges into
  reader-index machinery.

Loader: KIND-8 arms in load-reresolve-baked-reloc — baked hit patches
the record base; no-baked-record consumers mint (cell starts unbound,
so such sites use the fallback until an install arrives: graceful).

NOTE: the ADR-0017 cold-vs-warm .fco witness is NOT green — it was
already broken before this session (bisected: DIVERGED at cbb941f /
eaa0c57 / cb4275b; 7,469 differing bytes of producer-asymmetric
prelude-offset drift), invisible because test-fast-sharded excludes
the bundle tests.  Filed as cave#315 with the bisect data.  Stage A
is witness-NEUTRAL: +22 bytes over its base, and without these gates
the divergence was ~150KB of shape drift — the gates stay.

Verified: 8/8 sharded suite, smoke, fasl spike suite (16/0),
alexandria via the auto-loaded fasl, 12/12 package edge probes;
replay boot 160.66G vs 160.53G instructions (+0.13%, perf-neutral as
designed — the wins land at Stage C).  ADR-0032 status: Accepted,
Stage A implemented.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>