fix: type market-data sizes as Option<f64>

[wboayue]

Closes #716. PR-B of two, following #717.

The bug

Historical tick and histogram sizes were typed i32, but IBKR models market-data sizes as Decimal and ships them as protobuf optional string. A fractional wire value failed integer parsing and fell through unwrap_or_default():

// before
size: parse_i32(&t.size),   // wire "0.5" -> parse::<i32>() fails -> 0

That is real data loss on crypto and fractional-share feeds, and it is what the issue reports.

What changed

Type	Field	Before	After
TickMidpoint	size	i32	Option<f64>
TickLast	size	i32	Option<f64>
TickBidAsk	size_bid, size_ask	i32	Option<f64>
HistogramEntry	size	i32	Option<f64>
ContractDetails	min_size, size_increment, suggested_size_increment	f64	Option<f64>

All eight decode through parse_optional_decimal (added in #717): None means TWS sent no value — field absent, empty, or an "unset" sentinel — Some(0.0) is a real zero, and a malformed size surfaces as Error::Parse instead of decoding as 0.

The ContractDetails trio is included because contracts without size rules genuinely omit those fields, so the old 0.0 was indistinguishable from a real value, and a size_increment of 0.0 is nonsense rather than merely unlikely.

parse_i32 is deleted along with its last call sites.

Also in here

Allocation. The tick and histogram decoders build their Vec with an explicit capacity rather than collect::<Result<Vec<_>, _>>(), whose size_hint lower bound is 0 (the iterator may short-circuit) and so grows by doubling. This preserves the exact-capacity allocation the pre-change .collect() already had from an ExactSizeIterator — introducing ? is what would have lost it.

Fixture builders. The four historical fixture structs now hold the raw wire string rather than an i32, so fractional, sentinel, empty and absent sizes are all expressible. A size_wire setter keeps the edge cases to one line, and None omits the field entirely:

histogram_entry(125.5, 0.0).size_wire(Some("2147483647"))

Examples. Four async examples needed real changes, not just casts. histogram_data.rs's max_by_key(|e| e.size) no longer compiles (f64 isn't Ord) and becomes max_by + total_cmp; its bar-chart scale also had to stop being seeded with 1.0, which was a faithful port of .max().unwrap_or(1) but floors the scale and squashes every bar once sizes can be fractional. Option<f64> isn't Display, so display columns go through a small fmt_size helper that shows n/a rather than hiding a missing size behind a zero. That helper is duplicated across the four on purpose — each example must read and compile standalone.

Serialized shape. All five retyped structs derive Serialize/Deserialize and utoipa::ToSchema, so this also changes the JSON: a size is now 100.0 rather than 100, absent is null rather than 0, and the generated OpenAPI schema becomes a nullable number. That is the one break here with no compile-time signal, so it is called out in both the changelog and migration guide.

Tests

Regression tests pinning fractional sizes through all four decoders — these assert Some(0.5) where the old code produced 0. Plus malformed → Err per decoder, and ContractDetails absent-size-rules → None.

Sync and async end-to-end tests assert a malformed size fails the whole request through the public API. Sentinel and empty-string handling is not re-tested per decoder: parse_optional_decimal owns those semantics and covers them exhaustively in src/proto/decoders_tests.rs, so the domain tests only prove each decoder is wired to it.

Coverage on the touched modules: historical decoders 97.7%, contracts decoders 97.7%, proto/decoders.rs 84.9% (unchanged from #717).

Docs

CHANGELOG.md gains a Changed section and drops the "not yet covered" caveat #717 added. docs/migration-3.0.md §35 covers the field table, None semantics, and the three migration gotchas — the Ord break, integer accumulators, and {} formatting — with before/after snippets. The quick-migration checklist links to it. Grepped README.md, all docs/*.md and module rustdoc for the changed field names: no other references.

Full sweep green: cargo fmt, all three clippy configs, all three rustdoc configs, just test, cargo test --all-features, and both integration crates (build + clippy).

Known inconsistency

proto::HistoricalTickLast / HistoricalTickBidAsk are reused verbatim by the tick-by-tick path, so the identical wire field now surfaces as Option<f64> on the historical side (TickBidAsk.size_bid) and f64 on the realtime side (BidAsk.bid_size). The boundary is drawn by which struct the field lands in, not by wire semantics. The issue reported only the i32 truncation and the realtime types were already f64, so widening the break was out of scope — but it is a real seam, and the decimal quantity type below is where it should be resolved rather than by retyping realtime piecemeal.

Follow-up

Option<f64> is an intermediate. A dedicated decimal quantity type would let sizes round-trip the wire's decimal representation exactly instead of through binary floating point.

parse_decimal_or_zero's remaining call sites are not all provably safe — the C# client guards several of them with HasX ? StringToDecimal(..) : decimal.MaxValue, so upstream models them as "absent means unset" too. Bar::volume / Bar::wap are the clearest case: MIDPOINT, BID and ASK bars carry no volume, and EDecoderUtils.cs defaults them to decimal.MaxValue while defaulting open/high/low/close to 0 in the same function. Those are the next candidates; the helper's rustdoc names them.