Pine adapter: TradingView's margin, sizing and admission rules for one-position books - #330

TradingView sizes, admits and margin-calls a strategy's orders on its own
ten-significant-digit money. This change teaches the Pine adapter the rules
that synthetic TradingView tapes pin. Each rule sits behind its own switch in
MarginRuleSwitches (include/pineforge/source/pine_adapter.hpp), so a
regression can be bisected rule by rule. The switches are process-wide and are
not installed API; only tests change them. No kernel file changes.

Rules, and the switches that ship on:

• decimal_sizing: a default percent-of-equity quantity is the floor of the
  shortest decimal of sig10(E) / price, not of its binary quotient
  (6.279 rather than 6.2789).
• pooc_fee_sizing: under process_orders_on_close, a default percent market
  order is sized on its slipped execution price, grossed up by a percentage
  commission on the ten-digit grid. The grossed unit is computed in one
  function, tv_pooc_fee_grossed_unit.
• slipped_signal_admission: an explicit quantity opening from flat at full
  margin is judged at its signal close, on the money scale unslipped and on
  the price scale slipped.
• gain_loss_money: realized money is kept as two sums, gains and losses,
  folded into the broker-state hash. It stands in for the engine's equity
  only inside the regime the tapes pin: no fee anywhere (configured, in the
  run's fee model, or recorded at an opening), point value 1, no FX series,
  a power-of-ten lot step and at most one open lot. One predicate,
  gain_loss_regime, holds that regime. Elsewhere the engine's equity stands.
• dust_unit_call: a dust restore calls one unit.
• long_open_close_checks and negative_free_cash_call: a long's one-unit
  residual call at the open, the low and the close, compared exactly.
• coof_next_point_close: under calc_on_order_fills, a close placed after a
  call fills at the bar's next point.
• close_point_reversal: a close_all sized before the close point's call.

Shipped off: fill_price_recheck. It gates both the unified placement half for
default orders and reversals and the fill re-check. With it off, those orders
keep the earlier admission. The scraped population showed reversals and
openings that TradingView fills at a gap of several ticks and then
margin-calls, a regime no tape pins. The tapes that need the fill re-check
are listed in their tests and asserted with every rule on. Splitting the
switch is a precondition for ever turning it on.

Evidence and tests:

• tests/fixtures/margin_call_rules holds 266 single-position tapes: 199
  asserted, plus 67 recorded tapes of a short's call cut-off and slippage
  gate, whose rule is not pinned. It also holds 10 order controls replayed
  through handwritten hosts. tests/fixtures/margin_ledger_rules holds 381
  event-ledger tapes. TradingView's exports and the Pine sources are
  committed byte for byte.
• test_margin_call_rules_tapes and test_margin_ledger_rules_tapes replay
  every tape:
  • with the shipped switches: 170 of 199 and 269 of 381 match, and every
    unmatched tape is listed with its reason;
  • with every rule on: 172 and 373 match;
  • with each rule off alone: every switch costs at least one tape;
  • with every rule off: 133 match, the tapes the engine matched before.
• test_margin_rules_forward_replay replays a sample as a backtest and as a
  bar-by-bar stream, with every rule on, every rule off and the shipped
  switches. It requires byte-identical trades between the two modes on every
  split.
• test_publication_witness adds two checks:
  • nine fuzz run-passes where the G + L money once parted from the engine's
    equity are regression tests: their closed trades must equal, in every
    Trade field, those with gain_loss_money off;
  • no closed trade of the battery leaves its lot grid.

Re-pinned tests:

• tests/oracle/test_oracle_frozen_size.cpp and
  tests/test_default_qty_signal_freeze.cpp: 6.2789 -> 6.279. TradingView's
  own tape of the same quotient,
  tests/fixtures/margin_call_rules/sizing/sizing-decimal-floor-oracle-control,
  buys 6.279 in two byte-identical exports. tests/oracle.sha256 is re-pinned.
• tests/test_explicit_qty_fill_admission_l4b.cpp: the GREEN-G no-op moves to
  a funded capital of 10001, and a new assertion pins TradingView's drop at
  the original operands, citing boundary/long-admit-m1u and
  admit/admit-eth-s3-m07u.
• Four witness tables, re-harvested with each TU's own switch. Only hash
  fields move: quiet_bar 110 of 276 rows, publication 43 of 156,
  host_view_memo 13 of 13, and lookup_index's kFlip rows, reads and final.
  Every closed trade of every battery is unchanged in every Trade field.
• tests/test_adapter_sizing_relower.cpp guard (e) keeps the kernel's
  generic fee-divisor and sizing restatements forbidden in the adapter, with
  one explicit exception: the TradingView fee-grossed unit inside
  tv_pooc_fee_grossed_unit, which must exist once and cite its tapes.

Docs: docs/pine-adapter-kernel-notes.md section 1 states the rules, their
implemented scope, the switch table, the evidence and the known gaps. The
documentation anchors that the adapter's line shifts move are re-anchored.

Known gaps, listed in the tests:

• a carried short's call schedule, a short's call quantum at margin 50, and
  a multi-lot short's call at the high;
• a multi-lot long at margin 50 with a per-contract fee;
• the slipped long's call at the low;
• the residual call after a trade history on the G + L money;
• the money basis of a long's one-unit call after a history.

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

Measurement

Full population against the active baseline, after the exit-child-order merge: 7,989 probes, all measured, 7 up, 0 down, all strong → excellent:

• Two hard-lane ETH 15m probes (proozac98, Blackelon) and job-2550 on BTCUSDT 15m.
• vasudev on EURUSD 15m.
• antoniolinux, Willow and Blackelon on NYSE:F 15m.

One-rule-off measurements attribute every move to its rule. One NYSE:F probe stays within its tier, with a worse ladder rung. Replayed at the reference's rung, its trades are byte-identical: the move is a rung switch, not a regression. CI profiles are posted as pineforge/verify; the sweep is posted as pineforge/parity.

🤖 Generated with Claude Code