Harden take-profit discipline

[DarriEy]

Summary

Calculates take-profit decisions from token-scoped current inventory and
fee-net executable economics; distinguishes trailing stops; makes the
lifecycle and trailing bands configurable; adds schema-v50 decision telemetry
plus a read-only chronological calibration report.

Rebased onto main after #420 and #421. Until #420 the portfolio monitor's
exit loop raised on every tick, so no live exit ever executed and this policy
was being tuned against a dead path. It runs now, so every threshold here
moves money on the next tick. Six amendments followed, in descending order of
what they cost:

1. Entry-time fallback restored. Sourcing position age only from a
   token-scoped fills walk, and setting entry_dt = now on a miss, pins
   fraction_remaining at 1.0 on every tick — the target freezes in its
   widest band and the near-expiry band becomes unreachable. 37% of open
   positions have no fill ancestry (they predate the fills ledger, or are live
   rows a venue mirror created without writing a fill); 40% of those have a
   trades row the old path used. Resolution is now fills → trades → "just
   entered". The trades leg is book-scoped but not token-scoped, because
   trades has no token column. A size tolerance was added too: qty >= size
   is an exact float compare against a summed quantity.
   cost_basis was evaluated and rejected as the source — its only timestamp
   is updated_at, which records when inventory last moved, not when it was
   entered.
2. Entry fee dropped from the gate. broker/pnl.py realizes the exit fee
   only and builds the ledger event only on the SELL branch, so avg_cost is
   fee-exclusive and the entry fee is never booked. Charging it demanded a cost
   the books never record. Drag removed: 1.47 points median, 4.00 p90, 6.86
   max. The entry leg survives as a reported diagnostic.
3. HOLD telemetry off the money path. The write was awaited inline before
   exits.append(...), taking the shared serialized write lock once per
   evaluated position, every tick. It is now one executemany after the loop
   (>99% fewer lock acquisitions on a full book), with
   execution.exit_decision_retention_days (default 14) pruning via a bounded
   per-cycle delete. HOLD rows are kept as the calibration counterfactual.
4. Migration collision class closed. v50 is kept here; Add prospective maker observatory #411 renumbers to
   v51. A duplicate _migrate_vN_to_vM is silently shadowed by Python and
   ruff does not catch it — that has now happened twice. A new test AST-parses
   the class and asserts each step is declared once, dispatched once, resolves
   to a distinct function object, and forms an unbroken chain to
   SCHEMA_VERSION.
5. An unscoped read no longer defaults a live position to paper.
   mode_flag is None when settings.is_live is not a bool and the tracker
   holds no settings of its own, leaving the read unscoped over both books.
   (A real Settings.is_live ANDs three bools and so always is one — with the
   kill switch armed it is False, which scopes the read to paper rather than
   unscoping it — so this is the duck-typed path, not a live-trading one. It
   still must not hardcode a book: the wrong answer was silent.) Give a Position the book it belongs to so live exits can run again #420 gave
   Position an is_paper field, so the position is now asked directly.
6. Zero disables the trailing stop. Field(gt=0) crashed startup on 0.
   Relaxing to ge=0 alone would have been worse: unguarded, zero arms the
   most aggressive possible trailing stop rather than turning it off.

Stop-loss and trailing-stop band values are unchanged (byte-identical to
the originals). bot.py's exit loop and #421's readiness criterion are
untouched.

Verification

• 2304 passed, 5 skipped, 1 deselected
• Coverage 67.44% against the 65.0 floor
• ruff check . clean
• 17 mutations run against the six new behaviours, 17 killed — each
  behaviour was broken, the test confirmed failing, then restored

Migration

Upgrades schema v49 → v50, adding the exit_decisions telemetry table and a
position/time index.