hftbacktest 2.4.4 as the independent sim-lane cross-check engine: pinned install, upstream tests, L1 feasibility - #232

Installs a user-level Rust toolchain (rustup, rustc/cargo 1.98.1) and runs
hftbacktest's own #[test] suite at the pinned py-v2.4.4/rust-v0.9.4 tag:
`cargo test -p hftbacktest --no-default-features --features backtest --lib`
passes 22/22 (0 failed), covering the queue/depth-model unit tests this
cross-check's L1 conclusions rely on. The bare default-feature `cargo test`
still fails to build (the `live` feature needs libclang + C headers this
host lacks and has no sudo to install); recorded as a distinct, narrower,
unresolved gap rather than papered over.

Re-runs the L1 feasibility fixture (Scenarios A/B/C) under both
no_partial_fill_exchange and partial_fill_exchange: results are identical,
so the feasible-with-limits verdict is unchanged on this fixture (order
sizes never exceeded the touch's quoted quantity, so the two exchange
models could not diverge here -- recorded as an open gap, not claimed
covered). Adds a Lee-Ready aggressor-side inference step, presented as
proposed data preparation (not an hftbacktest feature), and shows it
restores trade-driven queue depletion on the same synthetic fixture.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…cker

HIGH: .github/osv-scanner-lockfiles.json was missing a covered_by_lockfile
entry for the new requirements.in (mover-v3 pattern), which failed
test_osv_lockfile_coverage.py. Added and re-verified.

MEDIUM, all grounded in a fresh read of the pinned Rust source:
- Corrected "8 files, 22 tests cover..." to the actual 7 files run under
  `-p hftbacktest` (connector/src/utils.rs is a different crate, excluded);
  explicitly states none of the 22 tests exercise exchange fill logic,
  IOC/FOK/GTX, ProbQueueModel, or the L2 depth-update path this fixture
  depends on, and that --lib skips 2 doctests (types.rs, latency.rs).
- Moved unrun claims (FOK, GTX, interpolated/per-symbol latency,
  FlatPerTradeFeeModel) out of "feasible": FOK and GTX are now actually run
  (Scenario G) and nonzero fees are now actually verified (Scenario H);
  interpolated/per-symbol latency and FlatPerTradeFeeModel are moved to
  "not exercised".
- Fixed the side-less-fill mechanism: depth alone (on_bid_qty_chg) never
  invokes is_filled() and never fills a resting order (Scenario D2, new,
  corrects an earlier wrong claim); passive fills need the OPPOSITE quote
  to cross (Scenario D3, new); trade-through fills are side-gated too, not
  just queue depletion (Scenario D1, new). Added the reviewer's three probes
  as tests.
- Renamed the mislabeled "Lee-Ready" rule to the correct EMO at-quote rule
  (Ellis, Michaely & O'Hara 2000; comparison at NBER w14158) and implemented
  genuine Lee-Ready (midpoint-based) separately; fixed the flat-tick
  fallback to reuse the last known direction; added a tape that actually
  exercises the tick-test branch and shows EMO/Lee-Ready can disagree.
  Documented the optimistic zero-queue-ahead bias for orders placed away
  from the touch (Scenario E, measured) and corrected the "exact
  single-level replacement" wording to price/size only, not queue position.
- Moved tests/ to the repository's top-level tests/ directory (the only
  blueprint that had its own tests/ subdirectory, which collided with the
  top-level tests package during unittest's bare discovery and so never ran
  in CI); made numpy a lazy import so the pure-Python EMO/tick-test
  classification tests run unconditionally instead of only "skipping
  cleanly".

LOW:
- A backtest-only, all-targets Rust build fails on six examples, not four;
  two (algo, custom_evhandling) are broken independent of the `live`
  feature or libclang, so installing libclang alone would not get a full
  `cargo test -p hftbacktest` to pass. Recorded the absence of
  rust-toolchain.toml/Cargo.lock upstream and this run's resolved versions
  (cargo tree, generated Cargo.lock hash).
- Added Scenario F: NoPartialFillExchange and PartialFillExchange genuinely
  diverge once order size exceeds the touch's quoted size; the cross-check
  plan now picks PartialFillExchange deliberately instead of calling the
  two equivalent.
- Relabeled the 69.2ms latency figure as the sensitivity point it is, not a
  calibration, and planned a sweep instead.
- Re-checked the maintenance precondition via the GitHub API: upstream's
  default branch is unchanged since the already-reviewed 2025-12-23 commit;
  py-v2.4.4 remains the latest tag.
- Reworded wheel provenance (PyPI-served digest match, not independent
  attestation); fixed the package count (43, not 42) and other nits.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…ions

Adds Scenario I: PartialFillExchange never depletes self.depth from a fill
(three back-to-back 100-share IOCs against an always-100 touch all fill in
full, position 300), matching upstream's own doc comment and the reviewer's
probe exactly. Corrects the cross-check plan to require matching (or an
explicit per-engine difference table for) liquidity/queue settings: reads
the actual sim-capacity branch (claude/sim-capacity-20260925, via git show,
not touched) confirming liquidity_consumption=True/queue_position=True and
the cited 180/900 fills-per-minute budgets, versus sim-paper-compare's
liquidity_consumption=False/queue_position=False; picks PartialFillExchange
deliberately as the closest-available (not equivalent) configuration.

Corrects the NautilusTrader taker-queue claim against the actual rc5
matching_engine/mod.rs source (snapshot_queue_position is only called on
the passive/maker branch, never for a taker fill) instead of an unverified
paraphrase. Fixes "coincide on A-D and G" to "A-D and G1/G3/G4" (G2, the
oversized FOK, genuinely diverges). Corrects the NBER 14158 citation
(Asquith, Oman & Safaya 2008, a short-sale misclassification study, not
"Diether, Lee & Werner" nor a literature review). Restricts the EMO rule to
exact-quote equality (was >=/<=), with new boundary tests confirming a
print beyond either quote now falls to the tick test. Actually runs EMO
inference on the D1 trade-through tape (with a seed trade so the tick test
has history) instead of asserting the result unrun.

Nits: FOK/IOC now imported from hftbacktest.order instead of hardcoded;
upstream publishes via maturin-action, not twine; the old test directory's
CI-invisibility is attributed to missing __init__.py under blueprints/,
verified with a minimal reproduction, not a name collision; algo.rs's build
failure is a Cargo examples/ packaging issue (a shared helper with no main,
also auto-discovered as its own example), not API drift like
custom_evhandling's; the "status understates the fill" caveat is scoped to
IOC only (FOK expires with exec_qty exactly 0); Scenario E is relabeled
hand-verified with its full trace, since ProbQueueModel's early-return path
never touches the probability formula here. Re-runs pip-audit with
--disable-pip --no-deps, now covering all 43 packages (was 42). Corrects
the stale "removal_commands_not_yet_run" cleanup record to reflect the
cleanup already executed, and records this round's own rebuild-then-delete
of the venv and gives rebuild commands in place of the deleted activate path.
Reframes the maintenance check as a dormancy risk (~9 months, 17 open
issues) with a concrete 2026-12-23 no-further-commits overturn checkpoint,
rather than reading as a pass.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

The cross-check plan and receipt had described the sim-capacity engine
configuration as of 0e0740e, which fed trade ticks into the engine. From
f47160b onward that runner feeds quotes only: in rc5, L1 trade ticks overwrite
both book sides, so 41% of IOC fills beat the NBBO when trades were fed.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…t; note the quotes-only change

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>