v0.18.0

The member crates move with the core because they now depend on nanobook 0.18
and its fractional Position.quantity: nanobook-broker 0.8.1 to 0.9.0,
nanobook-risk 0.6.4 to 0.7.0, nanobook-rebalancer 0.8.4 to 0.9.0. Their
published predecessors still pin nanobook 0.17.

Breaking (nanobook 0.17.2 → 0.18.0):

• Position.quantity is Shares, not i64.
• The Python Position.quantity getter returns a float.
• backtest_weights_with_options takes new optional fields on
  BacktestBridgeOptions; existing calls keep compiling, and every new default
  is a strict no-op.

Added

• Optional fill_schedule on backtest_weights / py_backtest_weights.
  Callers that decide on a coarse calendar (monthly rebalances) but want to
  fill on the real next trading bar can pass a parallel bar schedule. When
  supplied, NextBarOpen / NextBarTypical read fill_schedule[i] instead
  of price_schedule[i+1]; an empty entry means no fill for that period.
  When omitted, behaviour is unchanged. SignalBarClose ignores it.

Changed

• BREAKING: position quantities are fractional. Position.quantity becomes
  Shares, a newtype over i64 counting micro-shares (1e-6), and Portfolio
  gains quantity_step in the same unit — 1_000_000 for whole shares
  (default), 1_000 for 0.001, 1 for effectively continuous. Whole shares
  were not a rounding detail at small size: $1,000 spread over 20 names priced
  near $219 rounds every target to zero and the portfolio holds nothing. A
  newtype rather than an alias so the compiler flags every call site instead of
  letting them silently mean something 1e6 different; fixed point rather than
  f64 so results stay reproducible regardless of summation order. notional
  goes through i128 before narrowing, since micro-shares of a large position
  exceed i64. The order book is untouched — Quantity = u64 still drives
  matching, because a limit order book matches whole lots.
• BREAKING (Python): Position.quantity returns a float. It called
  Shares::whole(), so a position of 0.4 shares reported 0 and was invisible
  from Python. quantity_micro gives the exact integer and quantity_whole
  the old truncated value.
• The monthly value budget is an exact cap, not an approximation. Each
  order was tightened against the budget remaining at the start of a period,
  so every order looked compliant alone while the period overspent — three
  positions each fitting under a $20,000 budget spent $30,000 of it. It is now
  a running sum over the priority-sorted orders. This matters because
  transaction volume relative to portfolio size is a criterion that can
  reclassify a private investor as a professional trader for tax, and a cap
  that can be exceeded by half again is not a cap.
• CI now lints the whole workspace. cargo clippy ran without
  --workspace, so only the root crate was checked and broker, risk and
  rebalancer accumulated 141 unchecked findings. Clearing them was mostly
  mechanical: 46 were the 100_00 = $100.00 digit-grouping convention, which
  the root crate had declared crate-wide but the other two had not — now set
  once per crate in [lints.clippy], which reaches integration tests as well
  (each file under tests/ is its own crate root, so a lib.rs attribute
  cannot). Another 38 were machine-applied by cargo clippy --fix. The
  remainder were dead-code warnings on shared test mocks: mock_binance.rs
  and mock_tws.rs are each pulled into four test binaries, so a method one
  binary drives looks dead to the other three.
• Public API baselines regenerated (docs/public-api/) — they were eight
  releases out of date. Nothing was removed; the growth is scenarios and
  indicators arriving in v0.17.0, plus a rendering change in newer
  cargo-public-api. Regeneration is now a step in RELEASING.md, and
  docs/public-api/README.md records the tool version, since the output
  format is not stable across releases of it.
• rebalancer kill states plainly that it does nothing on Windows. It
  already refused there, but with a bare "not fully supported"; it now names
  the manual steps (taskkill, then cancel open orders at the broker) so an
  operator is not left assuming an emergency stop happened.
  docs/operations/kill-switch.md marks live execution on Windows unsupported.
• ibapi major bumps are now ignored by dependabot, with the reasoning in
  .github/dependabot.yml. 3.x removes the arm that detects IBKR disconnects
  mid-execution; it needs its own change driven by the mock-TWS drills rather
  than arriving as a weekly PR. Minor and patch bumps still come through.

Fixed

• Buys never spend cash the account does not have. execute_fill deducted
  an unaffordable buy with saturating_sub, so cash went negative and the
  portfolio held more than its equity — leverage financed by an overdraft no
  broker extends. On a $1,000 account (10 equal-weight names, US-equities-tiered
  costs) that stabilised around cash −$5.30 / invested 100.55%. Buys are now
  truncated to the largest quantity_step multiple whose notional plus
  CostModel::compute_cost fits in cash; a quantity that snaps to zero is
  skipped entirely (no zero-notional fill that would still charge
  min_commission). Sell quantities are never reduced, but a sell whose
  min_commission exceeds cash plus proceeds is skipped whole — reducing it
  would only make the floor worse. The buy-side trim is observable through
  trimmed_fill_count and trimmed_shortfall_cents so the portfolio cannot
  quietly diverge from the strategy with nothing recording why. Additive
  accessors only — new, with_quantity_step, rebalance_simple, and the
  execute_fill call sites keep their signatures.

Added

• Execution constraints on Portfolio, as parameters rather than policy:
  no_trade_band_bps (leave a position alone until it is off target by more
  than this), min_order_value (never place an order below this notional),
  max_trades_per_rebalance, max_order_value (truncates the order rather
  than dropping it, the way a desk works a large order down) and
  max_rebalance_notional. Which values are right depends on the account and
  the venue, so they are something to sweep or solve for, not to hardcode.

  These exist because of what fractional shares uncovered. A portfolio with
  constant target weights and flat prices placed 360 orders across 36
  rebalances of 10 positions; the correct answer is 10, the initial buy. The
  loop is driven by commission itself — a rebalance pays commission, equity
  falls, every target value falls with it, so next period every position reads
  as marginally over target and gets corrected, which costs more commission.
  Whole-share rounding used to mask this as an accidental no-trade band. A band
  of one basis point brings it back to 10 orders, which is the measure of how
  little those other 350 were correcting.

  Defaults (0, 0.0, None) are a strict no-op and reproduce the previous
  behaviour exactly. When a cap binds, the surviving orders are those furthest
  from target with ties broken by symbol, so the result never depends on hash
  iteration order.

• Windowed trade budgets on the backtest bridge: max_trades_per_day,
  max_trades_per_month, max_traded_value_per_month. They live in the bridge
  because Portfolio has no concept of dates and does not gain one. The bridge
  has no dates either, so the caller supplies period_day_ordinal and
  period_month_ordinal — plain monotonic integers. That keeps a calendar
  library out of this crate, since the only questions asked are "same day?" and
  "same month?", and it makes 24/7 venues fall out for free: several periods
  sharing a day ordinal simply share the daily budget. A budget is inert
  unless both the budget and its ordinal array are supplied — a cap can never
  be enforced against a calendar the caller did not provide, so absent ordinals
  fail open to no constraint rather than to a guessed one. Mismatched ordinal
  length is an error, not a silent truncation.

• last_rebalance_order_count and last_rebalance_notional report what a
  rebalance actually executed, after every filter. Additive accessors rather
  than a changed rebalance_simple signature, which is published API.

• quantity_step is settable from Python, so fractional sizing is
  reachable through the bindings; Position also exposes quantity_micro
  (exact) and quantity_whole (truncated).

• Frozen reference test for the native ChaCha20 Monte Carlo path
  (native_mc_matches_frozen_reference). The existing reproducibility test
  compares two runs inside one build, so it passes whether or not an RNG
  upgrade moves the stream — the rand 0.8 → 0.10 bump in 0.17.1 had to be
  checked by hand. The comparison is relative to 1e-12 rather than bitwise,
  because the terminal math goes through exp/ln whose last ulp is
  libm-dependent across platforms; any real stream change moves the numbers by
  whole percent.