Janady13/missionmatch / docs/evidence/phase2/functional/gate-a-phase2-close.txt

1GATE A, FUNCTIONAL. Phase 2 close, 29 July 2026. Re run after the scope completion work. 2 3Everything below was verified against missionmatch_staging on M4:5432 and the live staging API at 4missionmatch-staging.thatwebhostingguy.com, using the REAL Package A and Package B documents already stored 5there from the July trial. 6 7================================================================================ 81. GATE 2, THE DIRECTIVE'S OWN CRITERIA, ALL SIX NOW PROVEN 9================================================================================ 10 11 every extracted scope item carries a source citation schema enforced, ck_scope_items_cited 12 uncited model output is rejected, proven by test ValidateScopeCandidates, 6 tests 13 a candidate row fixture fails the scrub suite anywhere scrub/fixtures_test.go, start middle and end 14 a tenant document fixture fails it anywhere same file, and it is private by PROVENANCE 15 eight racing promotions elect exactly one requirement race n=8 and n=32, both winners=1 16 an ingest run during promotion corrupts nothing RAN FOR THE FIRST TIME, passes 17 approved requirements are never rewritten by ingest RAN FOR THE FIRST TIME, survived 20 rounds 18 byte for byte, md5 6acbad7d5a24cb161b141e1d5a99e717 19 20THE LAST TWO HAD NEVER BEEN EXECUTED. internal/integration/ingest_during_promotion_test.go sits behind 21//go:build integration and skips without pools, so the 53 package unit suite never touched it. It carried two 22of Gate 2's six clauses and its own positive control, and all three now run and pass. 23 24================================================================================ 252. PHASE 2 SCOPE THAT WAS MISSING, NOT MERELY UNVERIFIED 26================================================================================ 27 28Reading the directive against the code found four items that were never built. Each is now built and tested. 29 30 Section 5.1 TABLES listed among the required deterministic extraction targets. Nothing read 31 tables at all. ExtractTables plus ExtractRateRows added, structural rather 32 than prose based, with the prose gate left exactly as strict. 33 item 4 SPLIT and MERGE screen 4 names six review actions and four existed. 'merged' was a status 34 value nothing ever wrote and split had no representation. Migration 0084 35 adds the split status. 36 item 5 DRAFT GENERATION nothing derived a draft from reviewed scope, so the review to requirement 37 path had a hole in the middle of it. Generation plus manual completion. 38 Section 6 SCREENS 3, 4, 5 no UI existed for upload, scope review, or requirement building, and none 39 for the amendment view, while Gate C names every one of them. 40 41================================================================================ 423. THE REAL DOCUMENTS, AND WHAT THEY SAID ABOUT MY OWN WORK 43================================================================================ 44 45THE WAGE DETERMINATION IS NOT RECOVERABLE BY THIS BUILD, AND IT NOW SAYS SO. 46 47 before this session extraction_status done, 1 chunk stored 48 after extraction_status FAILED, no_text_layer, 0 chunks stored 49 50That is an improvement, not a regression. The one stored chunk was mojibake: this PDF's text layer decodes to 51character soup because its embedded font carries no usable character map. Reporting done over a garbage chunk 52is worse than failing, because a reviewer cites it and the pipeline's one promise is the citation. 53 54 MY TABLE READER MADE IT WORSE FIRST. Run against the real document it recovered "tables" of content like 55 "(0 CC" and "CC A3", because mojibake contains tabs and therefore satisfied every structural test. A 56 structural reader that bypasses the prose gate must carry its own quality bar or it becomes the hole in it. 57 Added: somewhere in a table block there must be a run of four letters containing a vowel. 58 59 THE SAME BAR CLOSED A PRE EXISTING HOLE. That mojibake span had been passing the PROSE gate all along. It 60 is now refused there too, which is why the document reports no_text_layer rather than done. 61 62 AND IT COST THE LEGITIMATE CASE NOTHING, measured on the same run: 63 04_PWS_original_10JUL26.pdf 54 chunks before, 55 after, GAINING one table of 3 rows 64 65 TABLE EXTRACTION IS PROVEN, just not on this document: 7 unit tests on a real shaped wage determination, and 66 the volume profile parses a 200 page PWS into 400 chunks, 200 tables and 1600 rate rows in 19ms. 67 68FINDING F6 FIXED AND VERIFIED ON THE REAL PACKAGE, where the first fix would have failed. 69 70 baseline 04_PWS_original_10JUL26.pdf (pws) 71 amendment 05_PWS_amended_23JUL26.pdf (amendment) 72 pairing cross_type, labelled, with the note that differences include differences between documents 73 result 1 MATERIAL CHANGE, quantity "1" to "7" 74 1 approved requirement derived from this opportunity, would flag Elevator Mechanic, PM 75 DRY RUN, nothing written 76 77 WHY THE FIRST FIX WOULD HAVE FAILED. Pairing purely within a document type looked correct against a fixture 78 I wrote where both PWS documents were typed pws. The intake actually classified the amended PWS as an 79 AMENDMENT, so pws held one document and the pairing fell straight through to the old wrong answer, the 80 solicitation against the amended PWS. It now matches the amendment to the baseline whose FILE NAME it most 81 resembles, with version words stripped. Previously this reported five changes that were partly artifacts of 82 comparing unlike documents; it now reports the one change that is real, and a headcount going from 1 to 7 is 83 the most important number on the page for a staffing product. 84 85================================================================================ 864. THE NEW CODE AGAINST A REAL DATABASE, BECAUSE OF F3 87================================================================================ 88 89F3 was the upload path being committed, reviewed and reported as built while never having written a row, 90because every test it had was pure. Split, merge and generation shipped with pure validation tests only, which 91is the same shape. Database touching tests were added and found TWO REAL BUGS within seconds: 92 93 MERGE USED FOR UPDATE WITH AN AGGREGATE. Postgres refuses it outright, "FOR UPDATE is not allowed with 94 aggregate functions", so the locking count was invalid SQL. The lock is the point, so the lock stayed and 95 the counting moved into Go. 96 97 A NIL GO SLICE IS NULL, NOT AN EMPTY ARRAY. required_certifications is text[] NOT NULL, so generation failed 98 on every document that named no certifications. The struct field looked correct either way. 99 100Both were invisible to pure tests. Six integration tests now cover split, merge, generation and manual 101completion, and they assert the properties that matter rather than the happy path: 102 103 split children INHERIT the parent's citation, quote hash and section label, parent retired to 'split' 104 splitting an already retired item is refused, which would otherwise orphan the first children 105 merge keeps the SURVIVOR's citation, absorbed rows retired to 'merged' pointing at the survivor 106 merging a nonexistent item is refused rather than silently retiring the survivor for nothing 107 generation from UNREVIEWED scope yields NOTHING, which is the human gate holding 108 generation is idempotent, second run created 0 skipped 1 109 a field the document never states arrives as unknown and is REPORTED missing, not guessed 110 111================================================================================ 1125. THE AUTHENTICATED LEG, END TO END ON STAGING 113================================================================================ 114 115New API sha 0fb3629d44d7b18a with the FIPS stamp, new bundle index-g1gTV6We.js. index.html was EDITED in 116place, only the two asset hashes rewritten, so all 22 lines and all 6 injected SEO lines survived, verified. 117 118 all four new screens serve 200 over HTTPS: documents, scope, requirements, amendments 119 the shipped bundle contains every new screen's strings, proven by grep against the served asset 120 121 A DEDICATED PROBE ACCOUNT, password set through cmd/rotate-credential so it goes through the real pepper 122 and PBKDF2 path rather than a hand written hash. Its tool grants were restored afterwards. 123 124 login 200, session 200 125 GET scope 200, GET requirement-drafts 200, GET amendments 200, GET documents 200 126 POST requirement-drafts/generate 200, {"created":0,"skipped_existing":0}, correct because Package B's scope 127 items are not in an accepted state 128 the amendment view returns the affected approved requirement, Elevator Mechanic, which is the same 129 requirement the diff independently flagged 130 131 TWO 403s AND A 404 WERE BOTH CORRECT AND WORTH RECORDING. The first 403 was the CSRF middleware refusing a 132 request with NO Origin header, which curl does not send: a hard Origin allowlist working as designed. The 133 404s that followed were the TOOL ENTITLEMENT gate failing closed on an account with no slateMATCH grant, 134 answering 404 rather than 403 so a surface is indistinguishable from a path that was never built. Neither 135 was a defect, and mistaking either for one would have sent me chasing the wrong thing. 136