Pin the frozen semantic corpus by its blob id, not a rewritten-away commit

[Parad0x-Labs]

What failed

tests/semantic_phase0/test_semantic_differential_eligibility.py — 3 cases red in the main
CI census (run 35570948370) and every PR run since; deterministic locally since the public
root:

• test_the_corpus_is_a_pinned_artifact_the_candidate_does_not_compute
• test_loader_refuses_a_trusted_object_that_disagrees_with_the_digest_pin
• test_candidate_files_and_import_shadows_have_zero_corpus_authority

Root error: git show c23257da…:ops/semantic_phase0_frozen_corpus.json exits 128 —
CorpusIntegrityError: the immutable frozen corpus object is unavailable.

Root cause

Same genus as the legacy-migration fixture (PR #21): the differential driver froze its turn
universe at a specific commit (c23257da…), a commit that no longer exists in the public
tree — the repository's history was rewritten at migration (public root 78f818b). The
loader correctly refuses closed, so the whole eligibility proof has been unavailable on
every public clone.

Repair — same content-addressed guarantee, rewrite-proof anchor

The corpus is now pinned by its blob object id
(c606c81b7b7b95abea94731c361037473df109a6) instead of <commit>:<path>:

• the loader still reads the trusted repository's object database, never the candidate
  worktree, never a worktree path, symlink, or import helper;
• a tampered corpus committed anywhere mints a different blob id — the pin refuses it
  (the original sabotage property is preserved);
• any commit carrying the exact artifact resolves the blob — the anchor survives future
  history rewrites that a commit-scoped pin cannot;
• the pinned sha256 digest and turn-count checks are unchanged.

The eligibility test's constant pin and its git show read are updated to the same blob-id
semantics (the test's sabotage assertions — AST scan for candidate replay_texts calls,
digest tampering, worktree/symlink/shadow attacks — are unchanged and still pass).

Validation

• Target suite: 3 failed → 30 passed (whole file).
• Cumulative: test_nonconsuming_policy_preview + the repaired file in one process:
  35 passed (the one failure is that file's own pre-existing CI-baseline case).
• Whole tests/semantic_phase0/ package: 630 passed, 4 failed = exactly the pre-existing
  replay-equivalence (2) / shadow-equivalence (1) / nonconsuming-preview (1) CI-baseline
  clusters — none in this PR's scope.
• ruff check . clean (pinned 0.16.7).

Base: main 8153a96. Two files changed: scripts/semantic_phase0_base_differential.py
(pin + loader read), tests/semantic_phase0/test_semantic_differential_eligibility.py
(constant + read, assertions unchanged).