…weep, prompt fidelity, evolution guardrails

Round 2 of the full-codebase audit: verified the 54-finding backlog the
round-1 verify fleet never reached (session limit) — 50 confirmed real, 4
refuted — plus 3 fresh finders over the interactions of today's three merged
PRs, which confirmed 11 more (7 major). One of the interaction majors was
CAUGHT LIVE during this session: the first morning on new code left NEE/CEG
naked because their entry limits were still working after the 15s protection
wait — exactly the hole fixed here.

ENTRY-ORDER LIFECYCLE (interaction majors)
- place_entry_protection treated "still working" like "terminal 0-fill" and
  walked away — a DAY entry could fill hours later with no stop watching.
  Now: still-working remainders are CANCELLED (protection-first; the
  post-cancel re-read protects whatever landed).
- No exit path cancelled a same-day resting entry BUY: a fully-exited symbol
  could re-open itself, and emergency liquidation could re-buy into the
  crash. Full-exit SELLs cancel the symbol's entry orders; emergency
  liquidation + force_delever cancel ALL entry orders first.
- Coverage repair read the last BUY under the executed-only predicate — a
  same-session 'submitted' BUY was invisible, so the belt no-op'd or applied
  a months-old stop level. get_symbol_last_buy(include_in_flight=True).
- run_intra_check now runs the coverage reconcile+repair every 30-min tick —
  the naked window for any late fill shrinks from hours to ≤30 min.
- get_current_stop_price reported whichever stop Alpaca listed first;
  multi-stop positions are now the steady state (per-BUY GTC stops). It now
  reports the HIGHEST (first to trigger) and logs multi-stop cases.
- Ex-div adjustment consolidated all stops into one level; now shifts EVERY
  stop down by the dividend, preserving per-lot geometry
  (broker.shift_stops_down with cancel/restore rollback).

BREAKER × SWEEP (interaction major)
- park_excess now refuses to park on a daily-loss-breach day (single choke
  point): the bookend used to buy ~99% of equity into SGOV right after an
  emergency liquidation, the next intra tick emergency-sold it with a
  spurious 🚨, repeating all day.
- force_delever unparks only what the deficit needs (not the whole T-bill
  balance) and records the vehicle's exit as SWEEP_SELL, preserving ledger
  isolation; recent-sells grading gets the symbol-level belt.

RISK / EXECUTION
- Non-finite cash failed the cash_only rule OPEN (NaN < 0 is False) — fails
  closed now, mirroring the NaN market_value guard.
- The R/R sanity recheck ran only when the ATR floor widened the stop; the
  raise-limit-to-market path (which shrinks reward the same way) dodged it.
  Hoisted: any execution-side geometry change re-checks R/R ≥ 1.2.
- ExecutionStage's post-review emergency liquidation priced limits off
  pre-LLM stale positions — refreshed first.

DATA / MEMORY / GRADING
- Earnings signal for missed-ops: structured Sentiment line (the naive
  "bearish" substring dropped NEUTRAL analyses), newest filing per symbol,
  90-day recency.
- One filing's failure aborted the whole earnings batch (no per-report
  isolation); prior-analysis pick was lexicographic (every 10-Q outranked
  every 10-K regardless of date).
- Calibration FIFO seeded BUY lots only from the lookback window, so a SELL
  closing an old lot matched an unrelated newer BUY; lots now seed from full
  history, the window bounds exits only.
- Missed-lessons "recurring" was auto-satisfied by one rolling-window streak;
  now ≥2 separated episodes OR ≥2 distinct symbols.
- Evening BUY grading compared a LIVE stock quote against YESTERDAY's SPY
  close (same-day buys always showed spy_pct=0); SPY leg is live now.
- get_ohlcv routed an all-NaN yfinance frame to [] without trying the Alpaca
  fallback; _staleness_days counted calendar days (every Monday read "3 days
  stale"); exDividendDate parsed in host TZ (+1 day drift on the SG host);
  correlation admitted 10-20-bar symbols as silent all-NaN matrix rows;
  close session's anti-flip-flop memory excluded the same day's midday row;
  a macro DB-write failure discarded a successful macro analysis;
  trail-cooldown wrongly ignored superseded (canceled) trail rows;
  fund_buys could report freed cash while ctx kept the stale pre-sale value.

PROMPT / AGENT FIDELITY (subagent batch)
- position_reviewer: entry stop/target lines no longer vanish for positions
  older than today; macro N/A fallbacks actually engage; grade section
  renders when only BUY grades exist; one malformed grade no longer destroys
  the whole EveningReport.
- risk_manager: prompt now carries equity/cash/per-position weights (the veto
  layer's rr_audit ran blind); allocation_pct labelled per action (BUY=% of
  portfolio, SELL=% of position).
- PM sees GROSS weights (was RAW — a "hold SQQQ" read as a 3x add);
  tech_analyst drops rows for symbols it never submitted; news_analyst
  renders mention symbols, gains a close-session mode, and word-boundary
  matches state-change symbols; meta_reflector renders watchlist_candidates.
- position_reviewer prompt no longer claims "TRAIL_STOP is always permitted"
  (the executor clamps it) or a universal SELL hard-trigger gate (it applies
  to already-trimmed symbols only).
- PM parse failure returns analysis_error (retryable, loud) instead of
  masquerading as a deliberate "no_trades" hold.

EVOLUTION / OPS (subagent batch)
- quarterly snapshot's budget `break` starved the Learnings section for
  every agent (break→continue; the hiding test now exercises the branch).
- apply-from-saved-reflection lane (EVOLUTION_APPLY_SAVED): what the human
  reviewed is exactly what gets applied — the documented gate re-ran the LLM.
- learning_text newlines defeated the line-based guardrails (normalized);
  empty apply runs leave an audit-log marker; auto-commit refuses to sweep
  operator-dirty prompt files.
- Evening/meta Telegram meta lines read the real ApplicationReport shape
  (were dead code); scheduler CronTriggers pinned to ET; healthcheck success
  pings skip intra_check (14/day pings kept the dead-man green while morning
  died); TELEGRAM_DISABLED accepts 1/true/yes/on in the wrapper.

Tests: 1430 green (+~110 new regression tests across 3 new test files).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PQkESoSTYx2bCy7WYnTPXR