test(cli): normalize F32 digits before the cross-version export comparison

[azimafroozeh]

What & why

Follow-up to #627, which turned the post-merge V5 ↔ V6 Format Fence job red on main.

• assert_export_fidelity compared the predecessor binary's export of the ml-intro embedding raw against the current binary's re-export: the v5 binary widens every F32 cell to 64-bit digits (0.10000000149011612), the writer from RFC 0051 prints the cell at its own width (0.1).
• The sibling helper canonical_export_rows already ran both sides through normalize_f32_and_nulls; the fidelity check now does the same before comparing, so it tests vector values, not digit spelling.
• The helper's doc comment names every predecessor binary instead of v0.9, since the same widening applies at the env-gated v3 and v4 seams.

Backing issue / RFC

• Trivial fast-lane: one test file, +7/-4, no engine change; repairs the fence assertion that feat: unify graph JSON output on one Arrow writer and offer Arrow IPC as a query result format #627 (RFC 0051, docs/rfcs/0051-json-output-via-arrow.md) invalidated.

Checklist

• Change is focused (one assertion in crossversion_upgrade.rs normalized through the existing helper)
• Tests added/updated for behavior changes (N/A: no behavior change; the fixed test is the coverage)
• Public docs updated if user-facing surface changed (N/A: test-only)
• Reviewed against docs/dev/invariants.md — no Hard Invariant weakened, no deny-list item hit (the check still requires the rebuilt vector to equal the original after both pass the one normalization)

Local verification

• cargo test --locked -p omnigraph-cli --test crossversion_upgrade current_v6_refuses_and_rebuilds_genuine_v5_and_v5_refuses_v6 -- --exact --nocapture with OMNIGRAPH_V5_BIN pointing at an omnigraph built from the pinned final-v5 commit 46b6d908 (the job's own invocation) — without this change it fails with the job's exact left/right (0.1 vs 0.10000000149011612); with it, 1 passed
• cargo clippy --locked -p omnigraph-cli --all-targets -- -D warnings -W clippy::dbg_macro — clean
• cargo fmt --all --check — clean
• v3 and v4 seams (OMNIGRAPH_OLD_BIN, OMNIGRAPH_PREVIOUS_BIN) — not run: no predecessor binaries built; same helper, same widening

Notes for reviewers

• The writer's spelling is the intended one (RFC 0051); the assertion was the wrong side. No release note.
• v5_v6_format_fence runs only when github.event_name != 'pull_request', so this PR's CI will not exercise the fence either; the local run above is the evidence, and the first CI run is the push to main.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.