feat: publish the export accounting valplay was recounting

[yakisoba0728]

Why

vrfkit already computes precise completeness accounting — chunks_processed, event_rows, movement_rows, event_layout_mismatches, error buckets, the nested net and sink objects — and quality_json_names_every_load_bearing_counter pins its keys.

The Python adapter emitted six manifest keys and dropped all of it. So valplay, the only consumer, derived every coverage and reconciliation verdict by recounting the NDJSON rather than reading the accounting this repo produces. The two could disagree silently, and its published quality model rested on the recount.

Nothing tested the seam either: renaming a column in vrf-export/src/schema.rs broke valplay while this suite stayed green.

What changed

A loss counter a consumer can actually gate on. The existing counters are unusable as gates — the reference replay reports skipped_bits 19,135,006 and 7,889 failed RPC streams while being healthy, so gating on a raw counter marks every match degraded forever. NetStats::lost_content_blocks() nets the preserved payloads back out, leaving only the lines docs/USAGE.md says must be zero. oracle and manifest now call the same function instead of restating the arithmetic.

The accounting crosses the seam. quality and net_field_export_groups pass through verbatim — no summarizing, no defaults. A missing quality writes null, not an empty object: zeros would read as "nothing was lost" when the truth is "nobody counted". The adapter's own measurements go in a separate adapter object so they can never be mistaken for upstream numbers.

players[] is deliberately not forwarded. valplay builds a richer version of the same table from the same rows, keeping character GUIDs as a set so revives and late joins keep their kill attribution. Forwarding a thinner copy would spread account PUUIDs into one more file and offer an attractive way to lose that. A test pins the omission.

Ordering is defined here. time_ms is not monotonic in packet_id — vrf-frame substitutes 0 for a non-finite frame time — so events are (packet_id, time_ms) stable-sorted and a regression is counted and published rather than sorted away, since sorting it away moves the stream off wire order.

The duplicated constants are pinned. The _ClassNetCache suffix and the unresolved-payload field name were literals in both languages with nothing tying them together. The test parses the adapter's assignment statements rather than grepping: both values also appear in its docstring, so a substring search would be satisfied by prose alone and could never fail. A third test feeds the scanner a decoy docstring to prove it reads the assignment.

The Swiftplay patch is applied. docs/swiftplay-metrics.patch described a fix to valplay that was never applied, so Swiftplay replays silently produced empty metrics. The change now lives in valplay and the patch file is removed.

Verification

cargo +1.86.0 test --workspace --locked      563 passed, 0 failed
tools/tests                                  481 passed
check_ascii.py --check                       117 tracked Rust files, ASCII only
clippy --all-targets --all-features -D warnings   exit 0
check_docs.py (full, not --fast)             exit 0
feature matrix, rustdoc, python_interop      all exit 0

Every new gate was mutation-tested: reverting each one turns the intended test red.

Deploy note

This must ship together with the valplay PR. The bundle schema goes to 2 and valplay's _validate_bundle requires it, so an old checkout on either side fails loudly by design. Existing bundles are invalid and the corpus needs re-extraction (~22s per replay for the adapter stage alone).

[yakisoba0728]

Downstream: yakisoba0728/valplay#4 — the consumer of this accounting. Ship together.