[green] reopen C3 post-OI NVS evidence at every stage

Turn the amended C3 NVS-receipt RED (51e9124) green per the v0.6.0
handoff's five-point robust design, without weakening any red:

1. Capture before evaluation: the verify-mode runner captures the raw
   NVS slice immediately after the full OI workload — after
   collect_observation, inside the still-open run resources, strictly
   before evaluate_thresholds and the output write. The read-only RUN
   probe (esp32.Partition, no NVS API import) streams the partition
   base64-encoded; the slice and acquisition log are written as
   exclusive mode-0600 files. The capture CLI pair
   --post-oi-nvs-slice/--post-oi-nvs-acquisition-log is C3 verify-only
   and required together.
2. Exact slice identity: validate_c3_post_oi_nvs_slice (bench) and the
   gate's private twin admit only offset 0x9000, size 0x6000, 24,576
   bytes, all 0xff, SHA-256 1df8949b...cdaba2.
3. Candidate binding: the C3 private result now embeds a post_oi_nvs
   binding whose summary carries the candidate release and firmware
   digests, the exact OI raw-log digest, 10 reset samples, the frozen
   geometry, slice and acquisition-log digests/lengths, integrity, and
   an empty written-namespace list; the receipt digest is recomputed
   from the canonical empty-inventory receipt, never copied.
4. Reopen-and-re-hash everywhere: create_result_file (creation),
   create_hil_completion_fragment (completion), and
   finalize_public_bundle (finalization) each reopen the exclusive
   mode-0600 receipt/slice/log trio — completion and finalization at
   the frozen sibling paths of the private result — re-derive every
   digest from the reopened bytes, re-verify the fully erased identity,
   and cross-check the retained OI raw-log digest and reset count
   against the bound observation via the new
   validate_c3_result_for_observation seam. Hand-written results and
   copied summaries are never trusted; both writers re-verify the
   evidence again before preserving output.
5. Diagnostics never mint: a capture that is not the exact erased
   slice is retained on disk for diagnosis but fails the run/writer;
   threshold evaluation stays upstream of any result write.

The FinalizationFixture's hand-built C3 result is brought up to the
amended frozen contract (embedded binding plus the sibling evidence
trio) so the real finalizer path keeps exercising it end to end. The
Pico writer/validator paths reject the new evidence inputs; public
summaries are unchanged and still carry no private receipt body.

Witnessed: test_v060_qualification_workflow 19 OK (was failures=5,
errors=6); test_release_v060_private_gates 18 OK (was failures=2,
errors=4); full harness 67 files, 10 failed — exactly the pre-existing
environmental set (missing firmware/upstream checkout, cwd-dependent
imports), with test_release_finalization 21 OK, waveshare gate 21 OK,
oi1 benches 78/15 OK, release contract 12 OK unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CAVghCmvKjF4BZFBUjAMwc
Signed-off-by: Viwat Vchirawongkwin <viwat.v@chula.ac.th>

Author: yyods