Add a journey — the standing onboarding pipeline

All paths relative to D:/Users/O185005/Documents/ERP/journey_v1 unless noted. Run python as python -X utf8 (Windows, UTF-8 JSONs). A journey is NOT part of the served corpus until step 7 is green and step 8 has run. Never leave the result uncommitted (the 2026-07-08 data-loss incident).

1 · Backbone

• Prefer _build/gen_journey_v1.py when the P7 pipeline inputs exist; else hand-build on the _build/gen_scminq_backbone.py pattern — entirely from the trace CSV ramco-erp-brain/berp-general/general/Process_Comp_Act_Task_services/<Comp>.csv (real activity/ILBO/task/service names; init/fetch/get/UI-cascade/help tool spine).
• EXTJS-pack slot inventory is MANDATORY when the component ships an EXTJS6 pack (berp-*/<bpc>/<Comp>/EXTJS6/<screen>.json — 560 components do): derive the slot set from the pack's input controls, not from SP parameters. The thin trace-CSV-only backbone is allowed ONLY when no pack exists, and the omission must be disclosed in provenance.gaps ("EXTJS pack not consulted/approximated"). The parity gate (step 7) enforces this.
• Must satisfy format.json (v4). Non-obvious requirements that bit before: post_options: [] is required; tool purpose and slot data_type/rule kind/bucket/basis kind must be in the format enums; on_fill is an ARRAY; post_conditions.output_variables entries are objects; every tool needs execution_path; thin_slot_view mirrors slots. Inquiry-class journeys: ALL filter slots obligation: optional (validator 7.2) — requiredness lives in rules.

2 · Depth mining (rules from raw SPs)

• Workflow{scriptPath: _build/depth_wave.js, args: ["<module>/<file>.json", ...]} (args must be a real JSON array), then bank: python -X utf8 _build/bank.py --output <task.output> --launched <targets> --tag <runid>.
• Every rule cites raw SP.sql:line; dead raises + err=None success-returns excluded.
• Additions are durable in _build/_additions/ — the recovery source of truth.
• Payload-leak tripwire (2026-07-10): agents sometimes serialize part of their return INTO the notes string (<parameter name="new_rules">[...]) and the top-level field arrives empty — bank.py WARNS on the pattern; when it fires run python -X utf8 _build/recover_leaked_payloads.py (report, then --apply + merge_depth). Never ignore it: 168 rules + 2 commit fixes were once silently lost.
• Resume protocol on limit stalls: _build/_WAVE_STATE.json (see memory jc2-depth-wave-resume). ONE workflow per reset window.

3 · The standing fixer chain — run after ANY merge/bank, IN ORDER

python -X utf8 _build/apply_commit_fix.py --additions <bank.json>  # NEVER skip — the lost-lane incident
python -X utf8 _build/normalize_recovery.py     # enum/tsv/untyped normalization
python -X utf8 _build/fix_template_deps.py      # FC-9 inferred-dep cleanup
python -X utf8 _build/fix_mismsg2.py            # (msg:) vs the P-XLSX error registry
python -X utf8 _build/fix_hygiene.py            # dedup + re-key + MANIFEST REGEN
python -X utf8 _build/fix_invariants.py         # class-vs-content (AFTER manifest)
python -X utf8 _build/prune_dangling_rule_slots.py

Trap: if a commit_fix proves the journey persists while classed inquiry-report, reclass to lifecycle BEFORE fix_invariants (else 7.2 strips the commit options — the 2026-07-08 circularity).

4 · Review-class conformance (the accumulated Ramco-review lanes)

python -X utf8 _build/review_july6_sweeps.py, then verify the invariants:

• Commit provenance: every commit option originates from THIS activity's own screens (the Express-PO class — foreign pipelines are navigation/linked).
• Labels = task truth: option/intent match Model_XML/trace-CSV task_desc (the PoEdtEntTrn2 Edit-vs-Delete class).
• No hidden commits: business-DML save-family tools (FC-3 oracle, vercel_deploy/brain_chatbot/data/sp_writes.json) are purpose: commit with an option (the Amend-Return class); fetch-page SPs with del/ret fragments are NOT.
• Lifecycle: carried display-only identity slot on the main screen; grounded discriminators only (entry filters = entry_search_filter); obligations follow SCREEN truth for fetched headers; fetch-first elicitation.
• Create-only leak: create-worker rules with only create* finish paths on a non-create journey get {path:<real>, enforced:false, reason} markers. Standing detector: _build/_lint_create_only_leak.json.

5 · Donor parity (only if the journey exists in ERP1)

python -X utf8 _build/reaudit_vs_erp1.py --assert-zero (whitelist W1–W8 + JC-6 verdict-archive routing). New w3_absent_jc6/w4_absent_jc6 counts = un-adjudicated donor knowledge → run a JC-6-style adjudication batch (_build/jc6_wave.js + _build/jc6_bank.py pattern).

6 · Enrichment parity

Target the corpus standard (reference shape: purchase/po_create_direct.json): activation gates, obligation_basis, valid_values, on_fill cascades, CW-7 tables_written evidence, CW-8 tax adjudication if the chain touches vat/tcal, E15 linked_journeys. Most arrive via the depth wave + overlays.

7 · Gates — ALL green before serving

python -X utf8 _build/verify_v4.py        # 0 SCHEMA / 0 REFERENTIAL / {} executor
python -X utf8 _build/validate_all.py     # only declared G-48 fails allowed
python -X utf8 _build/audit_corpus.py     # D1=0, FAB=0
python -X utf8 _build/reaudit_vs_erp1.py --assert-zero
python -X utf8 _build/check_journey_parity.py <module>/<file>.json   # PARITY GATE

The parity gate makes "as comprehensive as the corpus" pass/fail: P1 EXTJS-grounded slots (or disclosed), P2 depth-wave ran, P3 every rule grounded, P4 activation overlay when the register has rows, P5 obligation_basis on mandatories, P6 a real commit/output surface. A dimension passes by MEETING the bar or by an honest provenance.gaps disclosure — never silently. Calibration: 1,579/1,634 of the existing corpus passes (the 55 fails are the corpus's own tracked tail, not gate noise).

8 · Downstream + durability

1. In vercel_deploy/brain_chatbot: node --max-old-space-size=8192 scripts/build_index.js (default heap OOMs past ~2,300 journeys — the 2026-07-10 FS close; a crashed build leaves index.json written but build_manifest.json STALE — check built freshness) + node scripts/build_grep_index.js; bump CACHE_VERSION in src/answerCache.js; restart the server (kill node server.js workers, relaunch, health-check :5173/api/health); npm run check (journey-count test self-adjusts from the manifest).
2. Refresh the coverage tracker + report: python -X utf8 info_data/journey_status/build_journey_status.py (regenerates journey_v1/status.json — the done/pending activity tracker — plus the CSV/HTML coverage report; the new journey flips its activity to done).
3. Commit immediately (journey_v1 + chatbot data/ + status.json) and push; update _build/_WAVE_STATE.json. ALWAYS git add <paths> BEFORE a pathspec-scoped git commit -- <paths> — pathspec commits include tracked modifications but SILENTLY SKIP untracked new files (the 2026-07-10 FS incident: 78 new journey files sat uncommitted through two "wave commits"). Verify with git status --short -- journey_v1 | wc -l == 0 after committing.

9 · KG + journey-brain delta (once the KG is built)

Follow kg_plan/KG_PLAN.md §KG-NJ (NJ-1..NJ-8): corpus stamp → gen_kg.py re-run (node/edge delta; view journeys re-anchor capability rows, capability_id stable; alias map re-resolves) → KG-G-29 backlog row closed + versioned answer-key fixture flip → formula gates (golden fixtures never regress) → journey-brain/chia serving refresh → commit + push. KG scope law: the ENTIRE 13-module corpus, never P2P-scoped. Until the KG exists, steps 1–8 complete on their own.

Quick reference

backbone → depth-mine + bank → fixer chain (3) → review sweeps + invariants (4)
→ donor reaudit (5) → gates (7) → rebuild + restart + COMMIT + PUSH (8) → KG delta (9)