v6.10.0

Performance

request.security in backtests

Most of this release is about the cost of request.security() contexts on
historical bars. Output is unchanged; the per-bar path is still available
with PYNE_NO_SECURITY_BATCH=1.

• Same-timeframe contexts run as one batch round. A static,
  dependency-free, same-timeframe context without lookahead no longer
  handshakes with the chart on every bar. Its child runs the whole historical
  phase at once, and the chart pairs each bar's value out of the shared ring.
• lookahead_on higher-timeframe contexts are batched too. A same-symbol
  higher-timeframe context with lookahead=barmerge.lookahead_on used to cost
  one chart-to-child round trip per chart bar. It now replays the exact round
  sequence of the per-bar path (prefill, period closes and every developing
  re-tick) as one developing batch round. The plan is generated lazily from
  the chart's bar window: on a 30k-bar run the parent's peak RSS dropped from
  164 MB to 87 MB, and a 29783-bar window completes in 30 s against 42 s on
  the per-bar path. Modes that re-run main() within a candle
  (calc_on_order_fills, calc_on_every_history_tick, script timeframe,
  calc_bars_count, bar magnifier, live) keep the per-bar path.
• Contexts start lazily. A context's child process is spawned at its
  first read instead of its first signal, so a request.security() call in a
  branch the run never takes no longer costs a process and a round per bar.
  Producers a consumer depends on are started first.
• Scope-aware dependency analysis. The security transformer tracks
  locals per function, parameters per call site, returned aliases,
  nonlocal/global writes, flow-sensitive break/continue/return,
  read-only array/matrix/map functions, per-loop loop variables, and
  module-level field-only classes (Pine types). Spurious "every context
  depends on every earlier one" chains, which disabled batching and forced
  serial execution, are gone on the scripts where they appeared. On one
  multi-context corpus script the last round of this work alone cut the run
  time from 126 s to 9 s.
• Helper signals are lifted. A request.security() inside a helper
  function called once and unconditionally now signals from the top of
  main(), so its context starts in parallel with the others.
• Faster higher-timeframe bar-close walk. Session ends are prepared once
  per date instead of once per bar, which was the largest parent-side cost of
  many-context scripts. A 24h schedule's 23:59:59 end marker is now treated as
  midnight, so the last bar of the day is no longer confirmed one second early.

Fixes

request.security

• A runtime-resolved context (symbol or timeframe known only at runtime) that
  sits in a branch the run never takes no longer demands a data feed. This
  includes ternaries and and/or operands, which Pine evaluates every bar:
  a missing feed now logs a warning at resolution and raises the
  "No OHLCV data found" error only at the first real read of the context.
• A nested request.security() (one inside another's expression) is now a
  real dependency of the outer context. The outer child previously read the
  inner value as na even when data existed.
• A producer whose symbol or timeframe was still unresolved when its consumer
  started could leave the chart waiting forever; it now starts at its own
  first signal.
• The per-bar path could read the previous step's value on the first chart
  bar, because a second wake-up from the child answered the chart's next wait.
• A lifted wait placed in front of a helper's return made the chart wait for
  each child's whole bar.

Strategy engine

• Pyramid add under process_orders_on_close. An exit leg created by a
  pyramid add filling in the close pass now gets its fill attempt on the same
  bar instead of the next one.
• Price entry after a pending market reversal. A stop or limit
  strategy.entry placed after a market reversal on the same bar froze its
  reversal size at 0, sat out the flip bar and re-armed a bar late. The
  pending market entries are now counted. On the affected corpus strategy the
  trade match went from 0.9984 to 1.0.

Runtime

• barstate.isnew on history. It stayed false through the whole
  historical replay; it is now true on every historical bar, as on
  TradingView, so if barstate.isnew reset blocks of varip accumulators
  run.
• varip series history. Reading the history of a varip series raised
  'float' object is not subscriptable; it now works.
• na in and/or. and/or never yield na in Pine; a comparison
  with an na operand leaked out as the whole expression's value and a plot
  exported na where TradingView exports 0.
• Three-state bool comparisons. In a script that keeps Pine's three-state
  bool, a comparison with an na operand now answers na instead of false,
  as TradingView does (na == na included). Two-state (v6) scripts are
  unchanged.

Broker order sync

• A bracket for an entry that was never dispatched to the exchange is now
  deferred together with its entry while the position view is unconfirmed,
  instead of reaching the plugin as an exit for a position that does not
  exist.