Session clock: read each day's occurrences once per bar; time() arguments resolved without copies - #339

…ents resolved without copies

time(tf, session, tz) with a script's own session argument reads
TradingView's session clock outside UTC (clock_intraday_bar, #334). Its
scans asked arg_day_occurrences at every read: about nine cache lookups a
bar through arg_cache_day, for at most five days, and utc_zone at every
call. And every time() / time_close() call copied the session, its zone
and the calendar zone into local strings: two heap allocations a call for
a zone name one past the small-string buffer ("America/New_York").

- clock_intraday_bar asks each day's occurrences of the cache once per
  call and keeps the vector; a day whose ring slot already holds it
  resolved is read from the slot (resolved_occurrences), any other through
  arg_day_occurrences as before. A repeated read returned the same vector
  and changed nothing: resolving a day touches only its date and the dates
  next to it (arg_day_switch of a clock window's start and end), so the ten
  days one call can touch never share a slot of the 64-day ring; and each
  day is first read at the same point, in the same order, as before.
- utc_zone(tz) is asked once per session-cache entry (arg_cache_utc), at
  the same place in the same condition: it has no side effect and depends
  on the spelling alone (both zone-memo claim sites record
  normalizes_to_utc of the spelling).
- resolve_session_tz and the timeframe default name the call's own strings
  or constants by reference instead of copying them.

Same answer for every call, same cache and zone-memo history; no hashed
state, switch, input or API. One file, src/session_time.cpp: the language
calendar functions of ADR 0001's rule 1, called only by the Pine host and
generated code.

composite-vcp, 2 x 5,000 stream pushes (callgrind): 247.8 M instructions at
a80c623 (#336), 242.4 M at f3bf9f0 (before the session clock), 235.4 M here;
timeframe_time 2,512 -> 1,275 instructions a call. The other benchmark
strategies never reach this code (identical counts).

Values: a80c623's code run beside this one at every call over the whole
ctest suite (0 differences in 28,164 clock scans, 128,560 repeated reads,
3.2 M argument resolutions); a differential over generated sessions, zones
and timeframes, every switch of 2024-2026 walked minute by minute
(16.6 M queries, identical digests); the whole-suite per-trade dump (all 21
Trade fields at %.17g, every broker and stream state hash), the 30 witness
harvests (0 re-pins) and 307 tape and witness executables byte-identical;
corpus 312/312, moved 0; release ci_verify with the receipt rows
passes, 859/859 (a first run lost one timing-ratio row,
test_native_intrabar_lower_lookup, to parallel load; it passes alone in
both trees and in the re-run).

Timing, measured with a80c623 (#336) as the base; this commit carries
the same patch onto main, and the pull request re-times it there.
c6i.8xlarge (Xeon 8375C), GCC 13.3, CPU pinned, 4 independent
builds per tree, rotating paired blocks (61 ungathered, 41 gathered), every
process under exactly 22 environment variables; percent, positive =
slower; 95% intervals within 0.28 points (batch median), 0.12 (per call),
0.46 (p95); * = covers 0.

  percent vs                 native  ta-sma  bracket    udt    vcp
  batch median, ungathered (1 m input, 1 m script)
    14b05c9                 -4.22   -2.99   +0.35   -1.70  -2.61
    f3bf9f0                 -8.90   -3.26   +1.07   -4.67  -2.02
    a80c623 (#336)          -2.88   -1.02   +0.10   -2.10  -5.11
  per call (stream mean), ungathered
    14b05c9                 -3.94   -0.84   -0.31   -1.32  -1.06
    f3bf9f0                 -7.22   -2.00   -1.35   -2.93  +0.01*
    a80c623 (#336)          -1.24   -0.34   +0.37   -0.56  -2.58
  stream p95, ungathered
    f3bf9f0                 -7.28   -1.75   -1.26   -3.34  +3.19
    a80c623 (#336)          -2.32   -1.30   -0.09   -2.31  -0.17
  batch median, gathered (1 m input, 3 m script)
    14b05c9                 -6.02   -4.09   -3.19   -3.48  -3.84
    f3bf9f0                 -9.48   -4.82   -3.09   -5.80  -1.68
    a80c623 (#336)          -0.88   -2.03   +0.16   -0.84  -3.63
  per call, gathered
    f3bf9f0                 -6.69   -2.00   -3.77   -3.63  -0.67
    a80c623 (#336)          -0.08*  -0.14   +0.25   -0.35  -1.66

native, ta-sma, bracket and udt execute a80c623's exact instructions;
their differences against it are code placement. This file's code is 263
bytes smaller, and the linker places every object's unlikely section ahead
of all hot code, so nearly every hot function moved by 192-256 bytes. A
control build (not this commit) that pads the object back to a80c623's
section sizes puts the hot functions at its addresses: bracket
then reads +0.08* per call against it, and composite-vcp -9.1
batch median, -4.5 per call and -3.4 p95 (ungathered).

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