fix(engine): emit TV same-bar close+reverse sequencing under frozen default sizing - #146

…izing

Generalize the TV-pinned SHORT-seed default-FIFO close collision oracle
from its fixed-qty exact book to frozen PERCENT_OF_EQUITY / CASH default
sizing with unequal seed/entry quantities (finding 272; derivation
evidence/mc-timing-and-sequencing-derivation-20260812.json, closeReverseSeq,
fitted 25/25 exact on alpha-forge-liquidity-matrix-v2).

TV rule: with a SHORT seed qty S (entered on an earlier bar) and the exact
same-bar book entry(Long); entry(Short); close(Long)[no-op];
close(Short)[frozen target S], all filling at the next open P, TV emits
(1) the old short S exiting via order 'Long', (2) a zero-PnL dur-0 LONG
round trip qty L (frozen default qty) 'Long'->'Short', (3) a second
zero-PnL dur-0 LONG round trip qty min(S, L)
'Close entry(s) order Short'->'Short', and (4) an end-of-bar SHORT of
exactly max(0, L - S) under the final entry's id — the real opposite entry
is not queued; ordinary signal processing resumes from that position.

Engine changes (all five touch points in src/engine_fills.cpp):
- exact-book tagging: qty-type-aware seed gate (FIXED keeps the constant
  default==seed equality; PERCENT/CASH require the complete placement-frozen
  snapshot and matching frozen qtys on both entries); the projected
  final-short admission generalizes 2*seed.qty to L + min(S, L) and, for
  PERCENT (pct<=100), additionally mirrors the KI-54 frozen reversal
  re-check without its fill-time epsilon so a tagged book can never be
  half-declined mid-transaction.
- materialization live check + fill: the close order materializes its
  placement-frozen target CAPPED at the live long book, min(S, L).
- final-short live check: accepts lot2 == min(S, lot1) and
  position == lot1 + lot2.
- final-short kernel: closes both LONG lots and re-opens SHORT with the
  residual lot1 - lot2 = max(0, L - S) at the same fill price under the
  final short's id/incarnation/comment; flat when L <= S.

The FIXED cohort's pinned L == S collapses every relaxed gate to the old
equalities and residual 0 (byte-identical; test_short_seed_close_collision
unchanged). #143 dead-strategy.close, the declined-reversal close-leg
suppression, gap-reject, and pyramiding paths are untouched; non-collision
stale closes are still Removed. New unit coverage:
tests/test_short_seed_collision_percent.cpp (percent remnant L>S with
follow-up close of the remnant, percent flat L<=S, CASH remnant, gap-up
KI-54 decline non-trigger control, partial-close non-trigger control).
Full ctest: 125/125.

Co-Authored-By: Claude Fable 5 <[email protected]>