refactor: unify the three skip filters and make one-shot narrowing structural

[wboayue]

Closes the remaining one-shot/decoder follow-ups in plans/claude-md-knowledge-graph.md.

Two of the five items were already shipped in #736 (expect_type and the seven dispatchers) — the follow-up list itself had gone stale, which the plan now records as its own instance of the rot the migration was about.

Two commits: the conversion, then /simplify, which found real holes in the first one. Both are described below.

Item 1 — unify the third driver

TickDecoder::MESSAGE_TYPE becomes RESPONSE_MESSAGE_IDS: &'static [IncomingMessages], and tick.rs::classify filters with the same is_undeclared helper the sync and async subscription drivers use. One skip decision instead of three.

Caveat, recorded rather than papered over: collect_stream_decoder_impls matches impl StreamDecoder< only, so the three TickDecoder consts got the name but not the gate. Setting one to &[HistoricalTickBidAsk, UserInfo] passes all 1384 sync tests. The node says so out loud and the fix is a follow-up — TickDecoder::decode needs a backstop before the probe can distinguish "arm exists".

Items 2/3 residue + item 5 — expect_proto

request_helpers::expect_proto(expected, decode_proto) pairs the message type a one-shot asked for with the decoder for its payload, returning the impl Fn(&ResponseMessage) the helpers already take. A combinator rather than a fourth parameter: those helpers already carry six arguments against a budget of three.

Narrowing was opt-in — each domain hand-wrote a decode_*_message wrapper, and the sites without one decoded whatever arrived on the shared channel. The original follow-up put that at four sites; it is 26, over 13 decoders.

$ grep -rn "expect_proto(IncomingMessages::" --include=*.rs src/ | grep -v -E "_tests\.rs|/tests\.rs" | wc -l
50

20 replaced a wrapper, 26 replaced nothing, 4 replaced an inline if message.message_type() == .. guard. 30 decoder functions deleted.

The four option-computation sites the plan called "genuinely blocked" use a new fold_one_shot_mut that fold_one_shot delegates to, so the Some(Err) / None disposition stays one decision.

Item 4 — the last hand-rolled folds

head_timestamp, histogram_data, market_depth_exchanges, historical_schedule, sync and async. historical_data's fetch stays hand-rolled and should: it reads two frames.

Behaviour change. All four retry a connection reset at most 3 times instead of unboundedly — async head_timestamp recursed on a closed stream, the other three looped. Sync and async now also agree on what a closed stream means; they disagreed before. Changelog entry under ## [Unreleased].

What /simplify changed

Four review agents (reuse, simplification, efficiency, altitude). Efficiency found no regression — the combinator is a 4-byte stack value built once per call, not per retry — and confirmed two improvements (a Box::pin removed per retry, six unbounded spins capped). The other three found this:

The sweep had holes. WSH (4 sites) and next_valid_order_id (2) were never converted, so the "every one-shot narrows" claim was false while they existed. next_valid_order_id is the instructive one: it reads the shared RequestIds channel through fold_one_shot with a bare decoder. Adopting the fold helper reads as adopting the convention and isn't the same thing.

Nothing gated the pair. expected and decode are unrelated to the compiler, so a mispairing feeds one message's bytes to another's prost type. The altitude agent mutated three sites and only incidental round-trip tests failed. Added test_expect_proto_sites_match_the_roster (src/common/one_shot_pairing_tests.rs): every site scraped from src/, checked against a PAIRS roster in both directions, each pair required exactly twice so sync and async cannot drift. The same mutation now fails by name.

22 tests deleted. In the tests the first commit retargeted onto expect_proto, the decoder argument is never invoked — expect_type or require_proto fails first — so they asserted only the combinator's own behaviour and would pass with |_| Ok(()). Three canonical tests plus the roster replace them.

One revert, and it is the most useful thing in the PR. Three of the four agents independently said the expect_type inside the two single-type StreamDecoder::decode impls was a redundant third copy of RESPONSE_MESSAGE_IDS, quoting the new expect_proto doc comment. Removing them turned test_response_message_ids_match_decode_arms red within one run: a decoder with no match has no _ => arm, so expect_type is its backstop, and without it the probe reads the decoder as claiming an arm for every message type including Shutdown. The doc comment that invited the removal was the actual defect. Recorded on the node as a counter-example — convergent agreement across independent reviewers is not evidence; the gate is.

Docs. New node docs/rules/wire/one-shot-narrowing.md plus its index line — the convention had shipped as a doc comment on a pub(crate) fn, which just rules-check cannot see because it validates structure, not coverage. expect_type's error now names the expected type; it was the only runtime signal for a mispairing and withheld the identifying fact.

And a count I got wrong. The plan bullet cited grep -v _tests, which returns 56, not the 44 it claimed — four test files are named tests.rs with no underscore. Written inside the bullet arguing that a follow-up list is an ungated completeness claim. Corrected, with the correct filter.

Four fixtures that had been lying

Narrowing reads the message-type prefix, which no assertion previously depended on: MarketRule built with id 87 (it is 93), news providers led with the literal "newsProviders" (parses as no discriminant), market-depth-exchanges 71 (it is 80). All passed for as long as they existed.

Gate

cargo fmt; clippy ×3; rustdoc ×3; just test (all three legs); cargo build --examples ×2; both integration crates; just rules-check (32 nodes).

Deferred as restructuring

Widening the one-shot helpers to &mut to delete fold_one_shot_mut and give the option-computation sites retry; a ProtoPayload trait that would retire the PAIRS roster entirely; gating the three TickDecoder consts; deciding which one-shots should retry at all. All four are written up in plans/claude-md-knowledge-graph.md.