t27f: reel_gf16_nmse cells 15, 16, 24 cite the PUB-02 run instead of naming it missing #10023

specs/t27f/reel_gf16_nmse.t27 (the GF16 vs bfloat16 NMSE research reel, #10012) draws cells 15, 16 and 24 as gaps because "the sealed run PUB-02 asks for does not exist". PR #10021 (Closes #10019) landed that run: the harness specs/benchmarks/gf16_bfloat16_nmse_run.t27 and the manifest bench/results/nmse-railway-x86_64-0x2545F4914F6CDD1D-2026-10-17.json. The reel is now wrong in the other direction: it names as missing something the repo holds.

Boundary

• In: specs/t27f/reel_gf16_nmse.t27 only -- SOURCES, MISSING, the four measurement arrays, STATUS, the slide rows, the header comment, and the count tests.
• Out: topic_engine.t27 and topic_reel.t27 (no rule changes), the harness, the manifest, the protocol docs, cell 17 (the chip repo's manifest still says measured_silicon false) and cell 25 (no Delta_dB in the paper draft yet).
• Each status is whatever earned_status (via topic_reel.t27 statuses_earned) gives the cell's measurements. Cells 15, 16 and 24 are DELIVER steps; for the research domain DELIVER needs a review, and bench(nmse): first sealed GF16 vs bfloat16 NMSE manifest, from a t27 run harness (Closes #10019) #10021 has 0 reviews, so they earn DRAFT, not VERIFIED or PUBLISHED.
• The manifest is a software run on the Railway lab, not a silicon measurement; the reel says so and quotes no number that is not in the manifest.

User Scenarios

1. A reader opens the reel and sees steps 16, 17 and 25 as drafts that point at the harness and the manifest, instead of gaps that say no run exists.
2. A reviewer who approves bench(nmse): first sealed GF16 vs bfloat16 NMSE manifest, from a t27 run harness (Closes #10019) #10021's run can flip REVIEWED for those cells, and statuses_earned then requires them to be VERIFIED -- no other edit decides it.
3. Someone who sets TEST_PASS or LIVE on those cells without a review sees the test still demand DRAFT.

Requirements

• R1: cells 15, 16, 24 have a non-empty SOURCES entry naming specs/benchmarks/gf16_bfloat16_nmse_run.t27 and/or bench/results/nmse-railway-x86_64-0x2545F4914F6CDD1D-2026-10-17.json, and an empty MISSING.
• R2: GROUNDED true for those three; TEST_PASS, REVIEWED, LIVE as measured (0 reviews on bench(nmse): first sealed GF16 vs bfloat16 NMSE manifest, from a t27 run harness (Closes #10019) #10021; both artefact paths return 404 under https://t27.ai/t27/files/).
• R3: STATUS passes statuses_earned; count tests updated to the new published/verified/draft/gap totals.
• R4: a test asserts that a test pass and a live page do not lift those DELIVER cells, and that a review would.
• R5: slide rows reflect the new counts and the "software run, not silicon" boundary, and still pass slides_fit / stage_slides_bare.

Success Criteria

• SC1: counts are published 15, verified 1, draft 9, gap 2 (cells), and count_true(GROUNDED) is 25 (cells).
• SC2: t27c test-report on reel_gf16_nmse.t27 and topic_reel.t27 on the Railway lab: 0 failed tests, 0 vacuous.
• SC3: t27b test --check on the same two files on the Railway lab: 0 failed tests, with the same per-test verdicts as SC2 (verdict mismatches: 0).
• SC4: numbers quoted in the reel that do not appear in the manifest: 0.