report: opt-in replay extension (raw realized P&L, settlement markers, equity members, companion stream) - #607

native_execution_consumer.cpp reached ReportReplay members through the
forward declaration in engine.hpp; the two equity-recording call sites now
use the existing wrapper methods (same lines, no line shifts), which are
defined where the type is complete.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Adds a ctest row that runs run_json WITHOUT any replay switch on three fixed
scripts (a FIFO re-cut group, an ANY re-cut group, an ordinary close with a
report-only terminal row) and compares the default JSON report and the
pf-report-sink/v1 stream, as raw bytes, with digests recorded from the
engine before the extension (docker/replay_off_golden.json). The row fails if
any default-path byte changes and prints the digests it measured.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…-price vector

The default JSON and primary sink stream of the three fixed scripts now carry
the sha256 and byte count measured on main 95f4fd1 (record run rj-20261028t103035-87ee0b,
six FAIL lines against the empty golden, digests taken from them). The row
test_replay_off_golden is red if any default-path byte changes.

New fixture shape nondyadic: three separate unamended closes at decimal
tenths (100.7, 101.7, 102.9 entries) with a 0.07 per-contract commission and
one held lot. The oracle follows the engine operation order in plain binary64.
The test compares bits of each presented pnl, the left fold from +0.0 against
the replay net profit sum and every Q4, and shows that a 10-significant-digit
rounded fold differs, so the vector can tell rounded operands from exact ones.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…a mismatch

The first run on bbc3cb5 matched the three primary sink streams byte for
byte but not the three default JSON reports: same sizes, different digests.
The report names the build that produced it (engine.version_string, a git
describe, and engine.commit_sha), so a build of another commit can never share
those bytes. The check now replaces those two strings by a fixed token before
hashing; every other byte is compared as produced. On a mismatch the raw text
is printed so it can be diffed against the record. The default JSON digests
are re-recorded from main with the same masking.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

The previous mask looked for engine.version_string and engine.commit_sha in
the report. On main the run failed with "default report has no engine identity
block": the report holds the build identity only inside fingerprint.token
(base64 of the canonical provenance JSON) and fingerprint.digest (sha256 of
those bytes). Both are now replaced by a fixed token in the raw bytes, and the
decoded provenance with only the engine version_string and commit_sha blanked
is appended in canonical form, so the rest of the fingerprint is still
compared. Every other byte is compared as produced.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…he build identity masked

The default JSON digests are recomputed on main 95f4fd1 with the corrected
mask (fingerprint token and digest replaced, decoded provenance appended with
only the engine version_string and commit_sha blanked). The three primary
sink digests are unchanged and match the first head run byte for byte.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

… as well

The head run differed from main in the default JSON only by the short commit
in the report's plain fingerprint.provenance.engine block (commit_sha
95f4fd1 on main, the head's own short commit on the head); version_string and
everything else were equal, sizes equal. The check now also replaces the plain
engine version_string and commit_sha members (each must occur exactly once)
by a fixed token, besides the token and digest. All other bytes are compared
as produced.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…y build-identity member masked

Recorded on main 95f4fd1 with the check that masks the fingerprint token,
digest and the plain engine version_string and commit_sha members. The three
primary sink digests are unchanged.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…off-byte pin

The mask replaced the fingerprint token and digest by a fixed token and
compared only the decoded provenance, so a change of the token's spelling
(key order, separators, number spelling) or of the digest scheme could pass.
The check now also verifies that the digest is sha256 of the token bytes,
replaces only the two engine strings inside the token bytes, and compares the
sha256 of the resulting bytes and the digest length. The decimal-price test
comment now says what the assertion proves (at least one rounded P&L differs
and the folded total differs).

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…token bytes compared

Recorded with the check that also compares the fingerprint token bytes and
digest scheme. The same six digests were measured on main 95f4fd1 and on
main 1654497 (two runs), so the default report and the sink stream did not
move across those commits. The primary sink digests are unchanged.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Brings the branch onto the combined 1.8.0 batch. Conflicts resolved:

- src/engine_report.cpp, src/source/pine_strategy_host.cpp: the replay
  recording hooks, tagged pending mode and recut annotations are kept inside
  main's current streaming-metrics, retirement-capability and exit-group
  cache structure. The default path is unchanged.
- ci/cr150-source-bounds.yaml, ci/cr150-source-sites.json and the source
  reading files: main's rows and the replay branch's 18 groups and 43 sites
  are both kept; the replay identifiers are renumbered above main's and every
  moved source coordinate is re-read from the merged files.
- Twelve documentation files: main's plain wording is kept; the replay
  additions are restated in it and every citation points at the merged line.
  The ABI is main's version 5, unchanged by the extension.
- scripts/test_cr150_source_gate.py: the open-group count is 1483 groups of
  main plus the 18 groups of the replay branch, 1501 (the branch's own earlier
  commit changed the same pin from 1465 to 1483 for the same reason).
- tests/test_replay_abi.cpp pins ABI 5; the real-engine fixture states the
  module capability main's retirement fixtures state.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Main moved by one commit (the pin that the repository's own workflow files
pass its admission linter) after the previous merge. It changes only
docs/ci.md, scripts/ci_preflight.py and its test, so the merge is clean and
no citation of the replay extension moves.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…ified sink count

Main's fused sink path returns the trades and equity arrays as spooled
streams without a resident length. The replay join and the empty-array
normalisation asked len() of them and failed with "a streamed array has no
resident length" on the sink route. The length is now taken from the
verified sink count (T for trades, E for equity points) when the array has
no resident length; resident arrays keep len(). The default path is unchanged.

Also moves two source-line citations of the feature-parity page by one line
(exit_comment 1110, exit_from_bracket 1635) as the anchor check reports.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

… returns encoded columns

After the length fix the sink route still joined zero rows: main's fused
path returns the trades and equity arrays as already encoded JSON columns
that hold no rows to iterate, so the replay join found its R records exceeding
the JSON trades. When replay rows are requested, the arrays are now built
from the verified sink records (trades through the existing trade decoder,
equity points as dicts), exactly as the counted arrays of the earlier path.
The default path and the named-sink summary are unchanged.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>