Repository navigation
ci: build every registered Release target before merge - #881
luisleo526 wants to merge 37 commits into
Conversation
Exercise forty indicator lengths through actual order-fill recalculation, with independent values, ordinary execution and unaffected-indicator controls. Configure strategy settings through the public source-host configuration API.
Pair Kahan sources and compensated addends in one owning ring, retain Change history in an owning ring, and materialize the unbiased stdev deque only for that mode. Preserve arithmetic order and rewind behavior.
Exercise the real strategy exit path with one and two absent parents, native issuance counts, churn and run-reset controls. Preserve nonempty trading behavior across quantity, OCA and order-fill recalculation forms.
Read retained live fields from resident cohorts and cache a bounded set of immutable run-local cohort handles. Keep admission and authoritative history reads on cache misses, and clear derived handles at bind/reset.
The corpus grader compared trade counts as a net difference, so an unmatched engine trade and an unmatched TradingView trade cancelled out. count_abs_delta now adds every unmatched gated TradingView and engine trade, and excellent requires it to be zero. The net difference stays available as net_count_abs_delta; count_delta keeps the legacy relative difference for the strong tier. The pinned parity headline moves to excellent=301, strong=10, the regrade measured in #836. All 312 pinned strategy trade outputs still match their baseline hashes.
Bring the branch up to date with main before merging. No conflicts.
The eight exact-id cohort handles and cache cursor grow the adapter and enclosing strategy host by 328 bytes. The Linux x86_64 GNU Debug CI compiler reported PineExecutionAdapter=90720 and PineStrategyHost=109624 in build-debug run43723835831, test_abi_layout_probe.cpp diagnostics63/68. These replace the old x86_64-only pins90392/109296; no aarch64 value is inferred. The ordinary ABI probe records that architecture in focused tests.
Retire six stale deque groups, record four current groups as OPEN, and reclassify three mixed-shape groups as unreviewed. Refresh scoped site identities from the actual source declarations. No group is claimed bounded. Correct the exact OPEN union from1618 to1616 and its state ceilings.
Record both cache-id identities within their existing OPEN groups, retaining the combined1616-group inventory and its exact count assertion. No allocation bound or verifier exemption is claimed. Carry the cache-tail placement so older adapter member offsets retain their ABI pins. The size pins already on the parent remain unchanged. This applies the remaining correction to the exact combined parent5e039b4e, rather than adding separate issue-branch commits.
Fold the SHA256 pin for the Eigen fallback archive into the combined checkpoint and exit cohort fixes so one verification covers all three.
Tie the regrade totals to the population reported in #836 on 2026-11-11, without presenting them as a new measurement or an activated baseline. Keep the measured 312-reference parity expectation unchanged. Reword these two historical sentences to add their dated population: README.md: "The historical gross-count regrade recorded in [#836](#836) measured **7,959 excellent / 30 strong**: 24 probes moved from excellent to strong and none lower." CHANGELOG.md: "Corrected the previously published parity grading: counting unmatched trades net instead of gross overstated the excellent tier by 10 of 312 reference probes (311 instead of 301) and by 24 of 7,989 graded probes on the recorded population."
The pin is handled separately in #879. This restores the tree of the combined checkpoint and exit cohort fixes without it.
There was a problem hiding this comment.
Graphify reviewed this change.
Worth a look — the grounded gate found no coupling regressions or blocking issues, but 2 advisory finding(s) below merit a look before merge.
Not checked: tests were not run; formal verification proved 0 of 4 changed function(s) (1 sampled, not proven, 2 not verified).
Formal verification. 1 change(s) alter behavior, breaking input(s) attached. PR-changed functions: 2/4 verified (0 proven, 1 may-equivalent, 1 distinguished) · 2 not verified (2 vacuous).
Behavior changes: gke\_admit\_workflow\_findings changes behavior, here is the input that shows it.
The verifier found a concrete input on which gke\_admit\_workflow\_findings behaves differently before and after the change. If that change is intended, ship it; if not, this is your bug.
Guarantee: This difference was REPRODUCED, the verifier actually ran both versions on that input and saw them disagree. It is real, not an artifact.
Evidence: On input \{"workflow":"'a'","name":"''"\}, the old code produced \[': trigger must be exactly pull\_request\_target', ': trigger must restrict branches to main', ': trigger types are incomplete or reordered', ': run\-name must bind the PR number and head SHA', ':… but the new code produces \[': trigger must be exactly pull\_request\_target', ': trigger must restrict branches to main', ': trigger types are incomplete or reordered', ': run\-name must bind the PR number and head SHA', ':…. Paste that input straight into a regression test.
Not verified on this run: check\_commands (vacuous: never exercised), pipeline\_workflow\_findings (vacuous: never exercised).
Graphify review — findings
Adds an eight-way sharded Release build of every lab test target to both ci-post.yml and gke-pr.yml, compiling without running CTest and, on GKE, also for docs-only and .github-only PRs. Each shard records ctest -N and CMake/link inventories before and after a real build, build-tests-result runs ci_build_tests.py union to demand the exact union across all eight, and shard 1 first runs a broken-registered-header control. The build and result aggregate gates now fail unless both build-tests and build-tests-result report success, with no skip tolerance, and build_shard and release_targets own the per-shard build directory and CMake file-api reads that the union depends on.
Worth a look
- Concurrent shards can race on a shared build directory —
scripts/ci_build_tests.py:194· Escalate · medium- unverified hypothesis — survived the adversarial refute pass; needs human review
- Reused build directories can be rejected because of stale CMake file-api indexes —
scripts/ci_build_tests.py:62· Escalate · medium- unverified hypothesis — survived the adversarial refute pass; needs human review
Review partial — this diff was larger than one review pass covers, so later files were not reviewed; some findings may be missing.
Analysis details — impact, health, verification
Impact & health
Graphify review
Impact — 1112 functions depend on the 1015 functions this change touches.
Health — this change adds coupling hotspots:
- worse:
gke_admit_workflow_findings()— 18 callers, 15 callees - worse:
_job_values()— 33 callers, 3 callees - worse:
_job_block()— 20 callers, 3 callees - new:
release_build_workflow_findings()— 2 callers, 7 callees - worse:
pipeline_workflow_findings()— 1 callers, 13 callees
Verification — 1112 functions in the blast radius were not formally verified this run (proofs are advisory here).
Gate & verification
graphify gate
PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.
Advisory (not blocking):
- verification_scope: 1079 function(s) in the blast radius were not formally verified this run
Test selection
Test selection
1436 of 1436 test file(s) selected (100%) via static blast radius.
Escalated to a full run for safety (
noncode-change,uncovered-code) — the selection is not trustworthy on its own (see below). CI should run the whole suite.
benchmarks/tests/test_aggregate.py— full-run-safetybenchmarks/tests/test_check_provenance.py— full-run-safetybenchmarks/tests/test_compare.py— full-run-safetybenchmarks/tests/test_public_names.py— full-run-safetybenchmarks/tests/test_run_all.py— full-run-safetybenchmarks/tests/test_run_strategy_feeds.py— full-run-safetydocker/pineforge_sink_finalize_test.py— full-run-safetydocker/pineforge_sink_replay_rq_test.py— full-run-safetydocker/pineforge_sink_replay_test.py— full-run-safetydocker/pineforge_sink_to_json_test.py— full-run-safetydocker/quota_columns_test.py— full-run-safetydocker/quota_convert_test.py— full-run-safetydocker/quota_finalize_test.py— full-run-safetydocker/replay_rq_engine_test.py— full-run-safetydocker/run_json_artifact_window_test.py— full-run-safetydocker/run_json_bin_test.py— full-run-safetydocker/run_json_bounded_transport_test.py— full-run-safetydocker/run_json_bulk_equity_fuzz_test.py— full-run-safetydocker/run_json_bulk_equity_test.py— full-run-safetydocker/run_json_codes_test.py— full-run-safetydocker/run_json_converter_metadata_test.py— full-run-safetydocker/run_json_converter_session_test.py— full-run-safetydocker/run_json_diagnostics_test.py— full-run-safetydocker/run_json_equity_scan_test.py— full-run-safetydocker/run_json_fused_test.py— full-run-safetydocker/run_json_harnesses_test.py— full-run-safetydocker/run_json_metrics_only_test.py— full-run-safetydocker/run_json_misc2_test.py— full-run-safetydocker/run_json_misc_test.py— full-run-safetydocker/run_json_pfbar_test.py— full-run-safetydocker/run_json_provenance_test.py— full-run-safetydocker/run_json_quota_supervisor_test.py— full-run-safetydocker/run_json_replay_test.py— full-run-safetydocker/run_json_selected_sink_test.py— full-run-safetydocker/run_json_sink_lexical_test.py— full-run-safetydocker/run_json_sink_summary_test.py— full-run-safetydocker/run_json_sink_test.py— full-run-safetydocker/run_json_sink_vectors_test.py— full-run-safetydocker/run_json_spool_work_test.py— full-run-safetydocker/run_json_stringio_test.py— full-run-safetydocker/run_json_symbol_feeds_test.py— full-run-safetydocker/run_json_syminfo_test.py— full-run-safetydocker/run_json_trade_scan_test.py— full-run-safetydocker/run_json_window_budget_test.py— full-run-safetydocker/run_json_wrapper_work_test.py— full-run-safetyinclude/pineforge/native_run_spec.hpp— full-run-safetyscripts/derive_corpus_feeds_self_test.py— full-run-safetyscripts/fingerprint_self_test.py— full-run-safetyscripts/run_corpus_self_test.py— full-run-safetyscripts/test_abi_receipt_skips.py— full-run-safety- … and 1386 more
non-code file(s) changed (
.github/workflows/ci-post.yml,.github/workflows/gke-pr.yml,scripts/fixtures/ci_build_tests/CMakeLists.txt) → running the full suite for safety (a code graph can't see config/fixture/data deps)
changed code file(s) with no mapped test (
.github/workflows/ci-post.yml,scripts/fixtures/ci_build_tests/deferred.cpp,scripts/fixtures/ci_build_tests/fixture.hpp,scripts/fixtures/ci_build_tests/smoke.cpp,scripts/fixtures/ci_build_tests/support.cpp) — a coverage gap or a missing link — running the full suite rather than only the selected tests
Selection is safe under the controlled-regression assumption; always-run tests + a periodic full run are the backstops. Advisory — it never changes the check verdict.
Docs that may be stale (advisory)
AGENTS.md§ Rules for tests (lines 95-102): references changed symbolsnamesCHANGELOG.md§ Breaking for hosts (lines 923-962): references changed symbolscommandCHANGELOG.md§ 1.1.0 — 2026-10-04 (lines 3225-3439): references changed symbolsnamesCHANGELOG.md§ Reports and tools (lines 4122-4167): references changed symbolsnamesCLAUDE.md§ Rules for tests (lines 95-102): references changed symbolsnamesLEGAL.md§ Trademarks and affiliation (lines 59-64): references changed symbolsnamesREADME.md§ With an AI agent (MCP, Docker only) (lines 83-116): references changed symbolscommandREADME.md§ Building, testing and the gates (lines 474-520): references changed symbolscommandbenchmarks/README.md§ Reproduce (lines 82-145): references changed symbolsdiscoverbenchmarks/README.md§ Layout (lines 170-198): references changed symbolscommand,discover
…and 10 more.
Formal verification
Behavior changes: gke\_admit\_workflow\_findings changes behavior, here is the input that shows it.
The verifier found a concrete input on which gke\_admit\_workflow\_findings behaves differently before and after the change. If that change is intended, ship it; if not, this is your bug.
Guarantee: This difference was REPRODUCED, the verifier actually ran both versions on that input and saw them disagree. It is real, not an artifact.
Evidence: On input \{"workflow":"'a'","name":"''"\}, the old code produced \[': trigger must be exactly pull\_request\_target', ': trigger must restrict branches to main', ': trigger types are incomplete or reordered', ': run\-name must bind the PR number and head SHA', ':… but the new code produces \[': trigger must be exactly pull\_request\_target', ': trigger must restrict branches to main', ': trigger types are incomplete or reordered', ': run\-name must bind the PR number and head SHA', ':…. Paste that input straight into a regression test.
No difference found (not proven): No behavior difference found in \_sized (not a proof).
The verifier ran both versions of \_sized on many inputs and saw identical behavior every time. Strong evidence the change is safe, but evidence, not a proof.
Guarantee: Empirical: differential testing (both versions run on many generated inputs). A divergence on an untested input remains possible, so this is 'no counterexample found', not 'proven equivalent'.
Note: An input the sampler did not try could still differ.
Could not verify: Could not verify check\_commands.
The verifier did not have enough to check check\_commands, so it is saying so rather than guessing. No false assurance is the whole point.
Guarantee: No guarantee either way, this is an honest abstention, not a pass.
Note: Reason: not verifiable: all 90 sampled inputs raised on both versions — the function never executed, so 'no divergence' would be vacuous (mostly NameError — names the real obstacle, not a sampling gap)
Could not verify: Could not verify pipeline\_workflow\_findings.
The verifier did not have enough to check pipeline\_workflow\_findings, so it is saying so rather than guessing. No false assurance is the whole point.
Guarantee: No guarantee either way, this is an honest abstention, not a pass.
Note: Reason: not verifiable: all 63 sampled inputs raised on both versions — the function never executed, so 'no divergence' would be vacuous (mostly TypeError — names the real obstacle, not a sampling gap)
· 1 grounded finding(s) anchored inline below; 4 more finding(s) on lines outside this diff (see the check run).
| return findings | ||
|
|
||
|
|
||
| def release_build_workflow_findings(jobs: dict[str, str], *, pr: bool = False) -> list[str]: |
There was a problem hiding this comment.
release_build_workflow_findings()
fans out to 7 callees (efferent coupling).
Grounded coupling-delta finding (deterministic), not an LLM guess.
Relocate227 stale file:line citations to the current source windows. Only anchor line-number and range fields change. Prose, symbol claims, content pins and checkers remain unchanged.
Balanced unmatched trades can leave the two list sizes equal. Normalize the canonical gross unmatched count by the larger compared population in both aligned and zero-alignment paths. Strong grading and fragmented-FIFO metric eligibility share that percentage and retain their existing thresholds. The net count difference remains report-only. Add real-tape boundary regressions with below-limit controls, and correct the rescinded legacy-percentage expectations. Date the historical 24 movers and 7,959 excellent / 30 strong to the population reported on 2026-11-11. Keep the existing 312-strategy headline pin pending the actual regrade. Documentation provenance: intentionally replace or remove the following sentences. The earlier Excellent/Strong definitions are included for the net comparison with the preceding main revision. No other docs sentences are removed by this change. 1. README.md: "The historical gross-count regrade recorded in [#836](#836) measured **7,959 excellent / 30 strong**: 24 probes moved from excellent to strong and none lower." 2. README.md: "- **strong** — ≥ 95% matched, difference in gated trade-list sizes below 6% of the larger count, entries within 0.1% and exits within 0.5% at p90;" 3. README.md: "The separate `net_count_abs_delta` (`net`) reports the difference in list sizes." 4. README.md: "The percentage metric, `count_delta`, retains the legacy relative difference in gated list sizes for the strong tier and other percentage comparisons." 5. README.md: "It can be zero when both sides have unmatched trades; that does not satisfy the excellent count gate." 6. README.md (earlier wording in the net comparison): "- **excellent** — the same number of trades, ≥ 99% of TradingView's trades matched, the distinct-entry identity check passed, entry and exit prices within 0.01% and PnL within 1% at the 90th percentile (trailing-stop scripts are graded on the *production* profile: exits within 0.02% and PnL within 1%, since a trail fill depends on TradingView's sub-bar path);" 7. README.md (earlier wording in the net comparison): "- **strong** — ≥ 95% matched, trade count within 6%, entries within 0.1% and exits within 0.5% at p90;" 8. CHANGELOG.md: "Corrected the previously published parity grading: counting unmatched trades net instead of gross overstated the excellent tier by 10 of 312 reference probes (311 instead of 301) and by 24 of 7,989 graded probes on the recorded population." 9. CHANGELOG.md: "The corpus grader counts every unmatched gated trade on either side as a gross mismatch for the excellent gate." 10. CHANGELOG.md: "One engine-only trade and one TradingView-only trade count as two mismatches, so equal list sizes cannot satisfy that gate." 11. CHANGELOG.md: "The count percentage and strong-tier rules retain their legacy calculation from the difference in gated list sizes (#836)."
Waive both exact-id handle cache members with reconstruction evidence. The authoritative cohort id and immutable native handle are already folded at src/source/pine_state_hash.cpp:353-357. Admission is the only handle writer (src/source/pine_adapter.cpp:3298-3317); no cohort id is erased or given a replacement handle within a run. cohort_handle_cache_ copies those id/handle pairs. Exact spelling and a nonzero validity word gate every hit; misses read the authoritative table (src/source/pine_adapter.cpp:3335-3346). bind clears it at 1253-1257 and reset_for_run clears it with the table at 2669-2671. It cannot supply an independent order, position or trading decision through these paths. cohort_handle_cache_next_ selects only the next of eight eviction slots (src/source/pine_adapter.cpp:3322-3332). A miss still uses the same table; replacement history changes work, never the returned handle. Add test_cohort_handle_cache_rebuild: reconstruct a cohort checkpoint from the hashed records in insertion order, discard both cache members, and resume the same actual strategy. Compare applied fills, every trade field, per-bar hashes and final hashes, with fill recalculation on and off. Compare cold and warm lookups after real spill/restore for 17 ids, in both orders and all eight cursor positions, including exact-spelling misses. The test catches wrong-id hits, missing cold fallback, eviction-dependent answers, incomplete reconstruction and cache-dependent trading or hashes. The checkpoint covers the cohort component at a flat callback boundary; the other live host components stay in place. It is not a process image. The coverage checker and its admission rules are unchanged.
Print each failed sub-step's name and exit status, then list every failed binding stage in declaration order. Preserve raw child logs, exit codes, and skip and advisory behavior when later independent checks pass.
Forward positive finite quantity-step metadata through the JSON harness and live runner, preserving explicit-step precedence and minimum-only behavior. Cover both entry patterns through the real callers with retained value tapes and funding, metadata-order and gridless controls. Correct the obsolete symbol-lot-size statement introduced by 65c69b0 in docs/pages/run-failure-codes.md. The JSON harness now accepts positive finite mincontract and qty_step metadata and forwards it to the engine, so the old claim that the symbol's lot size is not accepted no longer describes the API.
Re-apply the one-line URL_HASH from #879. Its commit is already in this branch's history, but a later commit removed the line from the tree, so a merge cannot bring it back. The archive at the pinned URL hashes to this value.
The gross count percentage exposes an unchanged trade mismatch in validation/mtf-triple-tf-macd-hist-confluence-01. Keep its actual Moderate grade and allow only that full id at Moderate through aggregate reference verification until #883 closes. Every other strategy retains the Strong floor, and missing input or a lower tier still fails. Pin the already measured 2026-11-16 reference result to 301 Excellent, 9 Strong, 1 Moderate and 1 declared anomaly. Explain the expectation correction from 301/10/0 to 301/9/1 and keep the historical 2026-11-11 population and fact-rendered scoreboard distinct. Trades are unchanged; baseline promotion requires its own verification. Add CSV-driven tests of acceptance, rejection for other ids/categories, removal of only the policy entry, missing/lower tiers and single-strategy verification. The acceptance test uses the existing analyzer and aggregate main on both the base and candidate. Tests are authored and unrun. Refs: #836 #883 README.md result cell reworded in this change (verbatim original): "this repository's sweep of the v1.0.1 library: **311 excellent + 1 declared anomaly**; the <!-- pf:scoreboard.corpusProbes\|int -->309<!-- /pf --> of them in the measured population: **<!-- pf:scoreboard.scopes.corpus.excellent\|int -->309<!-- /pf --> excellent**" Documentation provenance for preceding changes in this PR (verbatim originals; these are earlier rewordings/deletions, not additional edits in this change): 1. README.md: "The historical gross-count regrade recorded in [#836](#836) measured **7,959 excellent / 30 strong**: 24 probes moved from excellent to strong and none lower." 2. README.md: "- **strong** — ≥ 95% matched, difference in gated trade-list sizes below 6% of the larger count, entries within 0.1% and exits within 0.5% at p90;" 3. README.md: "The separate `net_count_abs_delta` (`net`) reports the difference in list sizes." 4. README.md: "The percentage metric, `count_delta`, retains the legacy relative difference in gated list sizes for the strong tier and other percentage comparisons." 5. README.md: "It can be zero when both sides have unmatched trades; that does not satisfy the excellent count gate." 6. README.md: "- **excellent** — the same number of trades, ≥ 99% of TradingView's trades matched, the distinct-entry identity check passed, entry and exit prices within 0.01% and PnL within 1% at the 90th percentile (trailing-stop scripts are graded on the *production* profile: exits within 0.02% and PnL within 1%, since a trail fill depends on TradingView's sub-bar path);" 7. README.md: "- **strong** — ≥ 95% matched, trade count within 6%, entries within 0.1% and exits within 0.5% at p90;" 8. CHANGELOG.md: "Corrected the previously published parity grading: counting unmatched trades net instead of gross overstated the excellent tier by 10 of 312 reference probes (311 instead of 301) and by 24 of 7,989 graded probes on the recorded population." 9. CHANGELOG.md: "The corpus grader counts every unmatched gated trade on either side as a gross mismatch for the excellent gate." 10. CHANGELOG.md: "One engine-only trade and one TradingView-only trade count as two mismatches, so equal list sizes cannot satisfy that gate." 11. CHANGELOG.md: "The count percentage and strong-tier rules retain their legacy calculation from the difference in gated list sizes (#836)."
…in onto main Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brings in the current main so the change is verified against it. The change itself is unchanged; the only overlapping file, CHANGELOG.md, merges without conflict.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ne train Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Retain build-tests and build-tests-result in the independent corpus-toggle fixtures and their exact dependency diagnostics. Preserve the admission assertions, mutation cases, and single-match checks. Validation: complete corpus-admission module, 57 tests passed. Lane: ci-cr1152-corpus-fixture-1116-0401.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
| 'import sys; findings = sys.argv[2:]; ' | ||
| 'sys.stdout.buffer.write(f"checked {len(findings)} anchors\\r\\n".encode()); ' | ||
| 'sys.stderr.buffer.write(("\\n".join(findings) + ' | ||
| 'f"\\n{len(findings)} broken anchors").encode() + b"\\xff"); ' | ||
| 'sys.exit(int(sys.argv[1]))', |
|
|
||
| from verify_corpus import MATCH_WINDOW_SECONDS, analyze_strategy | ||
| from test_verify_corpus_trail_profile import TRADE_COLUMNS, TRAIL_STRATEGY | ||
| import verify_corpus |
Summary
Pull-request CI could stay green while a registered test target did not compile (main once carried a test that failed to build, and nothing failed). This makes the PR workflow build every registered Release target, sharded, and fail when any registered target is missing from the build or does not compile.
🤖 Generated with Claude Code