agent-progress.txt

# DeepResearchForecast agent progress

## Project overview

This repository implements a six-stage forecast workflow: DeerFlow 2 research, ontology generation, temporal graph construction, simulation preparation, OASIS simulation, and publication-gated reporting. The active feature strengthens the complete actor-evidence contract from deep research through the exact OASIS runtime persona.

## Session log

### 2026-08-16T17:35:00Z — branch `codex/shenzhen`

- **Feature/program:** `LOOP-017` measurable cost, reliable localhost, and forecast-quality refinement. The identifier was corrected after independent review because an older July 15 program already used LOOP-016.
- **Target behavior:** Make the six-stage workflow observably faster and more reliable while preserving provenance and Foglamp safety; route localhost directly to the DeepResearchForecast workflow; improve UI/accessibility, market integration, and decision visuals.
- **Read-only findings:** `pipe_f23527f7d903` has an 82,983,739-token durably recorded main-run meter total/lower bound, not a proven 162,733,517-token total; the larger number double-counts the synthetic 79,749,778-token research entry. True usage remains unknown because same-message-ID subagent usage is dropped and OASIS tokens are not recorded. Research is 96.1% of that recorded lower bound; graph is the time bottleneck at 8h37m, with twelve 900-second operations representing three aggregate timeout-hours. Current source already contains several historical research/graph mitigations that require validation rather than duplicate implementation.
- **Current-source defects:** Stage-1 usage drops later parent-plus-subagent cumulative snapshots; budgets do not govern RESEARCH/RUN subprocess spend; scenario-fork APIs reconstruct the wrong handoff directory; live ETA/heartbeat/spend helpers are disconnected; low-confidence Polymarket matches can change a forecast before provenance is removed; selected tool markets lack CLOB/history fields.
- **Concurrent work boundary:** Separate user-owned frontend and backend slices are present: the frontend routes `/` to ResearchView and improves styling/accessibility, while the backend adds SPA serving, fallback-model/fail-fast handling, visualization-axis/shared-Plotly changes, and pytest log isolation. The frontend rebrand still breaks the existing launcher readiness signature, its brand targets `/legacy`, and Vite plus `start.sh` can open two tabs. These external changes were inspected but not modified or attributed to this session; each is assigned to its corresponding future feature for review and verification.
- **Files touched by this session:** `PLANS.md`, `feature_list.json`, and `agent-progress.txt` only. The feature list was extended additively with nine implementation features plus one separately authorized controlled-run acceptance feature; the completed actor-intelligence entry was not altered.
- **Checks run:** Repository/memory/continuity inspection, three independent read-only audits, durable pipeline/log/artifact reconciliation, source tracing, and dirty-tree/concurrent-edit monitoring. No services, paid pipeline, provider call, runtime code, pipeline state, commit, or push was changed.
- **Result:** Audit and approval-gated plan complete; implementation pending user approval.
- **Next steps:** After approval, execute LOOP-017 Slice 1 only: reconcile and browser-validate the localhost/UI entry slice, including readiness signature, root/alias/history/unknown routing, single browser opener, responsive keyboard behavior, and no duplicate `/run`.

### 2026-08-16T18:03:34Z — branch `codex/shenzhen`

- **Feature/program:** LOOP-017 pre-implementation acceptance and concurrent-change re-baseline; no feature pass flag changed.
- **External changes reviewed:** Root ResearchView/router/SPA delivery and UI polish; fallback fail-fast/model guard; visualization axes/shared Plotly; pytest log quarantine. The external owner committed this combined 24-file slice as `3306af8` while validation was running. It remains user-owned, is not attributed to this session, and is not accepted merely because it is committed.
- **Launcher result:** The real `scripts/start.sh --detach --no-open` path reproduced the stale-brand failure after both services became healthy, exited 1, and safely removed its owned listeners/PID files. The current launcher test fixture is stale and false-greens. Vite and the shell also both own browser opening, Vite binds beyond loopback, and backend-port authority is inconsistent.
- **Browser result:** Flask-served `/`, `/research`, refresh, history, and unknown-route behavior worked with zero `/run` POSTs. Product-brand navigation incorrectly entered `/legacy`. Settings lacked modal semantics, focus entry/restore, an accessible close name, and Escape behavior. Root had no horizontal overflow at 320/390 px, but source review found mobile history overflow plus broader tab/progress/graph/dialog/contrast/reduced-motion gaps.
- **Backend re-baseline:** A concurrent owner landed serve-time Plotly bundle injection after the initial failing repro. Refreshed tests and a real Plotly HTTP replay proved the bundle is inlined under the opaque CSP and direct JS remains blocked; the stale Plotly delivery blocker was retracted. Graphiti still bypasses fallback-model guarding; cross-provider fallback can inherit the primary credential/base; and every HTTP 400 can poison the fallback tuple for 15 minutes. Exact `/api`, unknown non-GET API, missing asset, and missing-build response semantics also remain incorrect.
- **Checks run:** 107 initial focused backend tests; 31/31 frontend unit tests; production frontend build; scoped Ruff; `git diff --check`; actual launcher rollback acceptance; direct desktop/mobile browser route, focus, console, request, and overflow checks; 25 refreshed chart/axis/SPA/fallback tests; real Plotly HTTP/CSP integration replay. The app fixture kept 32 graphs and deleted zero. No provider or pipeline action occurred.
- **Files touched by this session:** `PLANS.md`, `handoff.md`, and `agent-progress.txt` only. This session did not stage, commit, or push; `3306af8` was created by the concurrent owner. Temporary browser artifacts were moved to Trash; ports 3000/5001 and managed PID files were clean at handoff.
- **Result:** External slices are partially functional but not accepted. LOOP-017 remains approval-gated; all ten new feature entries remain `passes: false`.
- **Next steps:** Obtain LOOP-017 approval, then implement Slice 1 red/green only: stable readiness marker, one optional browser opener, loopback/port authority, correct SPA/API/static fallbacks, root-safe brand, launch-intent idempotency, truthful preflight, and accessible/responsive modal/navigation behavior. Re-run the full launcher/browser matrix before starting Slice 2.

### 2026-07-21T19:56:45Z — branch `codex/foglamp`

- **Feature:** `feature_actor_intelligence_grounding_v2`
- **Target behavior:** Every simulation-relevant actor receives a distinct, source-grounded intelligence profile and a bounded actor-specific slice of global research context. Actor plans, incentives, investments, likely actions, alliances, competitors, constraints, and uncertainties also feed the unified deep-research report.
- **Initial state:** The repository already has an `actor-role/v1` safety and runtime-delivery contract, but the current default DeerFlow 2 research topology normally suppresses Track B and does not produce a deep actor dossier. The live implementation must be re-read before selecting the enrichment seam.
- **Files touched:** `feature_list.json`, `agent-progress.txt`, `init.sh`, `PLANS.md`, and `handoff.md` for the required harness and initial execution contract.
- **Checks run:** Repository/memory/continuity inspection only. Implementation and scenario checks are pending source reconciliation.
- **Result:** In progress.
- **Next steps:** Map the current actor producer, research manifest, role compiler, prepare artifacts, and OASIS consumer in parallel; decide the minimal versioned contract; add failing regression scenarios before implementation.

### 2026-07-21T20:12:45Z — branch `codex/foglamp`

- **Feature:** `feature_actor_intelligence_grounding_v2`
- **Milestone:** Producer/consumer/test source reconstruction complete; implementation contracts selected.
- **Findings:** Fresh default global-synthesis runs disable the dedicated actor Track B; structured extraction can promote any nonempty actor list; actor-role/v1 loses the requested forward-looking and decision-model fields; the full report never reaches the per-actor executable OASIS identity.
- **Decisions:** Run one shared baseline actor plane before global synthesis; carry `actor-intelligence/v1` inside the compatibility `actors.json`; materialize bounded `actor-context/v1` packs from exact report/actor bytes; compile `actor-role/v2`; bind the new packs into existing cast/profile/role/runner integrity checks.
- **Files touched:** `handoff.md` and `agent-progress.txt` only for this milestone; three code slices were assigned with non-overlapping ownership.
- **Operational recovery:** `uv cache clean` removed 49,713 disposable cache files (1.2 GiB) after a worker-spawn failure caused by a full data volume. No source or user artifact was deleted.
- **Checks run:** `./init.sh` passed; prior harness JSON/diff checks passed; implementation tests remain pending.
- **Result:** In progress.
- **Next steps:** Integrate the three implementation slices, add cross-boundary regression scenarios, update architecture documentation, synchronize the tracked DeerFlow runtime overlay, and run focused plus broader quality gates.

### 2026-07-21T20:23:15Z — branch `codex/foglamp`

- **Feature:** `feature_actor_intelligence_grounding_v2`
- **Milestone:** First-principles cross-slice contract review completed while implementation remained in progress.
- **Findings:** The research producer's stronger claim-level `dimensions` schema had diverged from flat-field consumer drafts; plan/action/investment qualifiers risked being flattened; generic intelligence tokens could contaminate report relevance; actor type drift could churn IDs; and Stage-2/graph compatibility summaries could omit the nested intelligence object.
- **Decisions:** Canonicalize the claim-level schema, retain flat aliases only for compatibility reads, preserve temporal/evidence/status qualifiers, separate public evidence from actor knowledge and analyst inference, require actor-specific relevance signals, make IDs stable across harmless type changes, and carry bounded intelligence into ontology/graph summaries.
- **Checks run:** Manual cross-file contract inspection of the in-progress producer, role compiler, context-pack builder, ontology handoff, and graph seed path. No executable gate was run against partially edited files.
- **Result:** In progress; concrete corrections were sent to each implementation owner before their contracts stabilized.
- **Next steps:** Receive and integrate completed slices, then run syntax/focused scenario gates before any documentation or diagram claims are finalized.

### 2026-07-21T20:39:43Z — branch `codex/foglamp`

- **Feature:** `feature_actor_intelligence_grounding_v2`
- **Milestone:** The PREPARE/runtime actor-context slice is implemented and has released file ownership.
- **Outcome:** Each selected actor receives a deterministic `actor-context/v1` pack bound to exact dossier/report fingerprints. Context is actor-specific, preserves public-evidence versus actor-knowledge boundaries, reaches both actual OASIS persona fields, participates in the role/profile/runner hash chain, and fails closed on new-contract tampering while keeping legacy `actor-role/v1` resumes compatible.
- **Checks run by slice owner:** 79 focused/regression tests passed, including 11 new actor-context tests; ruff on the new module/test and `git diff --check` passed. Root-level integration gates are still pending.
- **Result:** Runtime slice complete; whole feature remains in progress while research and role slices converge.
- **Next steps:** Review the released runtime diff with the other contracts, receive the remaining slices, then run the combined focused and cross-stage suite.

### 2026-07-21T20:54:26Z — branch `codex/foglamp`

- **Feature:** `feature_actor_intelligence_grounding_v2`
- **Milestone:** The producer, role compiler, PREPARE/runtime, synthesis-owner, Stage-2/graph, and tracked DeerFlow skill slices were integrated.
- **Independent audit:** Five release blockers were found despite the initial green focused suite: historical v1 roles were being recreated with a v2 compiler, partial v1 evidence could downgrade, returned packs could diverge from sealed bytes, inconsistent context/role counts could bypass validation, and raw context claims could reach the configuration LLM.
- **Response:** The runtime slice was reopened with one explicit negative regression per gap. The actor documentation and diagrams remained provisional until those gaps closed.
- **Checks:** The actor skill passed repository hygiene, deployed-overlay synchronization, and `quick_validate.py`; the first full backend run exposed four intentional mandatory-report-owner expectation changes plus one clean-source, actor-unrelated translation fixture failure.
- **Result:** In progress; audit verdict correctly remained not clean.

### 2026-07-21T21:06:14Z — branch `codex/foglamp`

- **Feature:** `feature_actor_intelligence_grounding_v2`
- **Milestone:** Qualifier round-trip and epistemic leakage regressions were added red-first and fixed.
- **Outcome:** Canonical nested plan/action/investment/decision/preference/capability qualifiers now survive repeated normalization; visibility is allowlisted; `actor_knows` is normalized only from recognized values. Analyst inference remains modeler-only, and contested/unknown evidence becomes actor-visible only with explicit boolean access while retaining its uncertainty label.
- **Checks:** The focused producer and epistemic tests pass; changed actor producer/compiler files pass Ruff; `init.sh` compiles all modified actor workflow modules and passes.
- **Result:** In progress pending the reopened runtime audit and final artifacts.

### 2026-07-21T21:21:30Z — branch `codex/foglamp`

- **Feature:** `feature_actor_intelligence_grounding_v2`
- **Milestone:** Runtime audit blockers closed; executable and visual validation completed.
- **Outcome:** Legacy v1 manifests now verify their original sealed fragments without v2 recompilation; nested v1 evidence cannot omit its top-level contract; returned packs/manifests are deep-decoded from exact persisted bytes; role provenance is checked against pack, pack-file, manifest, report, actors, dossier, source catalog, intelligence, and roster hashes; the runner rejects count/schema/platform/sidecar downgrades and validates every prepared platform; configuration projections recursively allowlist and sanitize context data. Blank top-level compatibility qualifiers also no longer shadow valid nested canonical qualifiers.
- **Tests:** 95 focused actor/runtime tests passed. The full backend suite produced exactly the known clean-source language-purity fixture failure; rerunning the entire suite with only that test excluded passed at 100%.
- **Visual checks:** Both editable tldraw scenes regenerated and validated through the official browser export path: whole system 198 shapes / 89 arrows / 178 bindings at 7080x4766; DeerFlow 2 subsystem 150 / 66 / 132 at 6380x4445. Native-resolution inspection found no clipping or topology error. The compact actor-provenance visual also passed desktop/mobile rendering and interaction checks.
- **Result:** In progress only while the final prose, JSON inventories, and independent settled-tree audit finish.
- **Next steps:** Reconcile documentation/inventory counts, rerun source/link/JSON/lint/hygiene gates, accept or resolve the independent audit, then mark the feature passing and close `handoff.md`.

### 2026-07-21T21:30:51Z — branch `codex/foglamp`

- **Feature:** `feature_actor_intelligence_grounding_v2`
- **Milestone:** Final broad producer-to-runtime audit superseded the provisional closure and reopened implementation.
- **Reproduced blockers:** Actor extraction can complete without a sealed roster; post-seal reconciliation can mutate the cast; source-less claims can be promoted through an evidence-gap waiver; raw research strings can reach ontology/configuration model prompts; unqualified or string-like knowledge access can become actor-known; the dossier ledger is not roster-bound; an all-gap dossier can pass without a judge; and fetched-content receipts/hashes/length are dropped before sealing.
- **Response:** Assigned three non-overlapping red/green remediation slices for producer/provenance gates, immutable sealed-cast orchestration, and prompt/epistemic isolation. Added a deterministic behavior-ready floor across five evidence families for every Tier-1/2 actor.
- **Checks:** Independent audit supplied concrete reproductions for all eight failures. The prior 95-test and visual gates remain useful evidence but are not acceptance evidence for these newly exposed paths.
- **Result:** In progress; documentation and diagrams remain provisional until combined regressions and a fresh independent settled-tree audit pass.
- **Next steps:** Integrate the three fixes, close any remaining Graphiti research-data boundary, rerun all focused and full backend gates, reconcile architecture artifacts, and obtain a clean independent re-audit before marking the feature passing.

### 2026-07-21T22:02:52Z — branch `codex/foglamp`

- **Feature:** `feature_actor_intelligence_grounding_v2`
- **Milestone:** Exact OASIS v1 delivery and the first stable cross-stage integration gate passed, then a deeper semantic Stage-1 audit deliberately reopened release acceptance.
- **Completed since the prior entry:** Actor-enabled extraction and parent reception now fail closed; sealed v1 casts are immutable; canonical roles/context use only source-bound claims; ontology, graph, and configuration model inputs are sanitized and delimited; exact Reddit `persona` and Twitter `user_char` are role-only v2 bytes; unsupported explicit future schemas cannot downgrade to legacy flat behavior.
- **Checks:** The stable producer-to-runner selection passed 410 tests. Three extra future-schema regressions passed. Targeted Ruff passed with only the profile/config files' documented pre-existing `F541`, `B905`, and `C420` categories excluded.
- **New adversarial findings:** Twelve offline reproductions exposed ambiguous/untiered cast admission, source-less relationship split-brain, unrelated-URL and stale-plan laundering, roster-only dossier binding, Stage-1 prompt reinjection, truncated/malformed judge attestation, oversized late-actor omission, stale resume lineage, homonym/alias collapse, evidence-free gaps, Track-A receipt leakage into Track B, and a duplicate helper overriding actor claim parsing.
- **Response:** Two non-overlapping implementation slices now own producer/provenance/identity/lineage and model-boundary/judge/report-coverage remedies. No full-suite or documentation result will be treated as final until all twelve have red/green regressions.
- **Result:** In progress; no paid research, OASIS run, provider mutation, publication, commit, or push occurred.
- **Next steps:** Integrate and independently review all twelve fixes, rerun semantic and combined gates, then execute full backend/hygiene gates and reconcile every architecture artifact before closure.

### 2026-07-22T02:01:06Z — branch `codex/foglamp`

- **Feature:** `feature_actor_intelligence_grounding_v2`
- **Milestone:** Complete. The first-principles actor-intelligence contract is sealed end to end from shared Track-B research through report, ontology, graph, PREPARE, exact OASIS runtime fields, final simulation-config seal, runner admission, and direct-child byte use.
- **Implementation outcome:** Every Tier-1/2 simulation actor has an exact semantic identity and 17-dimension source-bound intelligence profile; unsupported claims become typed exhausted gaps rather than invented detail. Plans, investments, incentives, actions, capabilities, constraints, relationships, decision rules, likely actions, and red lines feed both the global report and a distinct bounded actor context. Current-v1 runtime behavior comes only from deterministic `actor-role/v2`; legacy flat fields, private/modeler-only evidence, raw report/graph prose, synthetic Reddit demographics, and generic fallbacks cannot override it.
- **Integrity outcome:** Producer, dossier, report, roster, claims, relationships, lineage, context, role/profile, cast, config, graph seed, and direct-child load are hash/identity bound. Authorized scenario/WorldState config changes are idempotent and resealed; completed reuse validates without rewriting and rediscovers adjacent current-v2 evidence; genuine unsealed v1 compatibility remains isolated.
- **Documentation outcome:** English and Mandarin READMEs, the actor-intelligence report, whole-system atlas, DeerFlow 2 atlas, approach comparison, 95-flow/101-route inventory, 100-family LLM census, and both editable/rendered tldraw scenes describe the settled architecture in parity.
- **Tests:** 1,037 selected cross-stage actor tests passed. Final collection is 2,815; the counted full run is 2,803 passed, 11 expected xfailed, and one unchanged actor-unrelated language-purity fixture deselected. Seven calendar tests that temporarily hit the intentional 2.0 GiB disk floor passed 7/7 after deleting only this session's pytest temp trees.
- **Artifact/quality gates:** Exact 101/101 Flask AST route parity; 400 local links; 1,131 source line ranges; 13/13 bilingual architecture-link parity; Pandoc over six documents; four JSON inventories; official tldraw validation at 198/89/178 and 150/66/132; `init.sh`; compileall; actor skill validation; changed-file Ruff with only documented legacy classes excluded; strict DeerFlow producer Ruff; `git diff --check`; no Vite/Chrome/preview debris. The final independent behavior and citation audits returned CLEAN.
- **Known unrelated issue:** `test_language_purity_batches_large_residual_set` remains red because its unchanged `_BatchLLM` lacks `.chat`; neither the fixture nor `report_agent.py` was modified by this feature.
- **Safety:** No live/paid research, OASIS simulation, provider change, external publication, generated `deer-flow/` mutation, commit, stage, or push occurred. `deerflow_bridge/.cache/` and unrelated dirty work remain preserved.
- **Next steps:** None for this feature. A separately authorized live run can validate external provider/search/OASIS behavior later; it is not required for the completed offline implementation gate.

### 2026-08-17T18:05:00Z — branch `codex/shenzhen` (created by a concurrent session at fcf7378; same tip as codex/foglamp)

- **Feature:** continuous improvement loop, iteration 1 (study + first fix slice)
- **Study:** 7-agent audit over code + run logs. Tokens: research = 96% of metered (79.75M at 49:1 in:out, full-thread resends ×3 tracks); graph + sim are metering blind spots (contextvars lost in executor workers; sim meter never persisted). Time: graph = 62% (8h37m, retry storms, 60% chunks discarded re-deriving structured actors.json content); sim re-runs vs exhausted quota and report publish-gate rework account for the rest.
- **Landed (commit 3306af8):** '/' → Research UI (+lazy legacy, tokens.css, brand, i18n sync, ConfirmDialog, SVG icons); Flask serves dist as SPA at :5001; fallback-model 400 bug fixed + invalid-model cooldown; dual-outage fail-fast; metric_trajectories axis fix; plotly 'directory' mode + serve-time bundle inlining (CSP sandbox preserved); pytest log quarantine.
- **Tests:** 25 new focused green; visualizer 191, research-wiring 55, viz-manifest/pdf/SPA 41 all pass; full suite: only the 1 known pre-existing language-purity failure. Frontend build exit 0, 31/31 unit tests; real-dist smoke check green.
- **Next:** iteration 2 targets by rank — graph-stage ingest bounding + retry caps (~8h), research thread compaction/caching (tens of M tokens), metering blind-spot fix, run-level provider halt, report section concurrency + language-purity cap, Polymarket transport swap. NOTE: several targets live in files carrying other sessions' uncommitted work (graph_builder, pipeline_orchestrator) — re-check git state at wake before editing.

### 2026-08-16T18:06:17Z — primary validation reconciliation after concurrent commit `3306af8`

- **Scope of correction:** The preceding concurrent-owner entry records useful implementation and test evidence, but its word “Landed” MUST NOT be read as feature acceptance. This primary-agent entry preserves the commit while recording the user-scenario failures that its focused/full suites did not exercise.
- **Current tip:** `3306af8` contains the 24-file UI/SPA/fallback/visualization/log slice. The refreshed Plotly serve-time injection is green and the root SPA is directly loadable, but the real launcher still rejects the current HTML and rolls back. The product brand enters `/legacy`; Vite and the shell can open two tabs; Vite exposes the loopback-trusted API beyond loopback; port authority is split; exact/non-GET API and missing-asset/build fallbacks are incorrect; preflight can false-green; launch retry has no server idempotency key; and modal/keyboard/contrast/mobile-history acceptance remains red.
- **Fallback correction:** The generic fallback guard does not protect Graphiti's direct fallback construction. Offline probes also prove cross-provider primary credential/base inheritance and provider-wide cooldown from request-specific HTTP 400s. These are release blockers for the fallback portion of the commit.
- **Cost correction:** Historical graph attribution loss is not proof of a current ContextVar defect because current source already propagates context into Graphiti executor work. The durable token denominator remains incomplete for different reasons: same-message-ID Stage-1 subagent deltas and RUN subprocess tokens. The accepted baseline remains the lower-bound accounting in LOOP-017, not a reconstructed exact 150M total.
- **Status/next:** All LOOP-017 flags remain false. Runtime correction still waits for explicit LOOP-017 approval, then begins with Slice 1 red/green only; Slice 2 cannot start until the launcher/browser/idempotency/accessibility acceptance matrix passes.

### 2026-08-16T18:27:21Z — branch `codex/shenzhen`

- **Feature:** LOOP-017 read-only forensic baseline and plan refinement; no runtime feature was started.
- **Milestone:** Four independent evidence streams were reconciled across RESEARCH, GRAPH, RUN/REPORT, and Polymarket/visual delivery. The study used the source-backed codebase-research workflow and preserved current-source versus historical-run boundaries.
- **Token result:** RESEARCH is 79,749,778 tokens and 96.1% of the recorded main-run lower bound, but totals are not complete. Failed lanes can disappear: `pipe_0f2bee7bd649` retains 25,161,961 attempted tokens while state reports 6,110,831. Reference adaptive passes consumed 20,286,821 tokens, including six zero-convergence Track-1 passes costing 15,817,428.
- **Time result:** GRAPH is the dominant reference critical path at 8h37m. Its sequential outer loop incurred twelve whole-batch 900-second deadlines, exactly three critical-path hours, and caller-accounted 278/466 chunks as skipped. The required fix is per-episode salvage/deadlines plus structured-v1 parity, not a blind concurrency increase.
- **Quality result:** REPORT prompt retransmission and duplicate graph payloads are a measurable secondary cost; interrupted RUN checkpoints are written but disabled for default pipeline resume; historical ensemble reports spent 6,356,626 tokens for zero aggregate agreement. Polymarket has a P0 influence-boundary defect plus reversed-token and concurrent-ledger P1s. Plotly delivery has a sibling-symlink/size containment P1.
- **Files touched:** `PLANS.md`, `handoff.md`, and `agent-progress.txt` only. `feature_list.json` remains unchanged in this milestone and all LOOP-017 flags remain false.
- **Offline verification evidence:** Graph owners reported 116 focused tests plus one custom telemetry bridge fixture green. Market/visual owners reported 417 focused tests green, while four new deterministic boundary fixtures reproduced defects not covered by that suite. Research/report artifacts were reconciled read-only; no paid/provider call was issued.
- **Independent QA:** Recomputed arithmetic and authorization gates passed. The reviewer found two missing traceability checks, so LOOP-017 now explicitly tests lost-response launch idempotency/local-only API/static failure behavior and provider credential/cooldown isolation. The 299 graph-seed count was clarified as persisted operations rather than the sum of actor/relationship/alias categories.
- **Concurrent drift:** An uncommitted external `prediction_markets.py` transport edit appeared after the audit. Read-only reconciliation found only browser-UA, bounded retry/backoff/timeout, and transport-diagnostic changes; it does not close the forecast-influence, CLOB identity, or resolution-ledger defects and was left untouched.
- **Concurrent telemetry test:** An untracked external `test_telemetry_fallback_attribution.py` appeared during final verification. It covers inferred LLMMeter attribution with one active run, not F1 provider credentials/cooldowns, and its historical lost-ContextVar premise conflicts with the current full-path propagation fixture; it was preserved for later source reconciliation.
- **Snapshot freeze:** The final status also gained concurrent `report_agent.py`, `telemetry.py`, and `test_prediction_markets.py` edits after the fan-in. They were not reviewed as settled implementation. The evidence snapshot is frozen at 2026-08-16T18:35:53Z, and affected later slices must rebaseline before editing.
- **Safety:** No runtime edit, pipeline start/resume, provider mutation, service launch, publication, staging, commit, or push. All unrelated dirty actor-intelligence and documentation work remains preserved.
- **Result/next:** Evidence baseline complete; implementation is gated. Obtain explicit approval of LOOP-017, then work only Slice 1 through launcher/browser/idempotency/accessibility acceptance. A paid comparison run remains separately gated.

### 2026-08-17T18:50:00Z — branch `codex/shenzhen`

- **Feature:** continuous improvement loop, iteration 2 (commit b83b716)
- **Landed:** telemetry single-active-run fallback attribution (graph tokens now metered + budget-enforceable; unattributed_process persisted; per-test registry isolation in conftest after the full suite exposed cross-test leakage); language-purity escalation guard/cap/short-circuit (outage worst case 180→3 strong-tier calls; known failing test now PASSES — zero standing failures); Polymarket backend transport hardened (browser UA, backoff, error-class diagnostics) + surgical uncommitted fix to deerflow bridge _polymarket_get (UA, 8s/2 attempts, transport_error_classes persisted).
- **Notable:** section-concurrency audit finding was stale (default already 6) — comments corrected, no flip; Polymarket config knobs deliberately NOT added (bridge reads env directly; unwired Config mirror = dead code).
- **Tests:** full backend suite exit 0 (first fully-green full suite on record here); 500+ focused across telemetry/report/market suites; ruff clean.
- **Next:** iteration 3 — dataviz quality package on clean files (uncertainty encodings, near-empty chart slots→tables, GraphPanel visual encoding, Step4 chart gallery); the graph-stage 8h fix + research 78M-token fix still gated on the dirty-file/checkpoint decision (concurrent codex session active on this checkout).

### 2026-08-17T19:42:00Z — branch `codex/shenzhen`

- **Feature:** continuous improvement loop, iteration 3 (commit 63c1954)
- **Landed:** honest uncertainty on forecast charts (self-consistency spread → p_low/p_high with provenance, never fabricated; point estimates labeled as such); chart density gates with markdown-table fallback + manifest skip reasons; PNG label wrapping and collision fixes; GraphPanel visual encoding (degree sizing, arrowheads, temporal dim/filter, search, stable CVD-validated palette — legacy palette failed the validator); live simulation trajectory chart + contained /trajectory endpoint.
- **Tests:** full backend suite exit 0 / zero failures; 291 viz + 26 new + 10 endpoint + 44 frontend; fixtures visually QA'd. Gate lesson: one run mis-collected from repo root (445 collection errors) — always gate from backend/.
- **Next:** iteration 4 decision point — the remaining top-ranked items (graph-stage 8h re-derivation, research 78M-token resends, run-level provider halt, orchestrator research-telemetry leak, _SLIM_EDGE_KEYS temporal fields) all live in files carrying prior-session uncommitted work; plan is a labeled checkpoint commit of that work (precedent 801faf4) unless the concurrent codex session shows activity.

### 2026-08-17T20:42:00Z — branch `codex/shenzhen`

- **Feature:** continuous improvement loop, iteration 4 (commits a53d9ed checkpoint + fd3e077)
- **Checkpoint:** prior-session actor-intelligence work (77 files) landed as labeled snapshot per 801faf4 precedent; secrets canary clean; unblocked the hot files.
- **Landed:** graph-stage cast-relevance chunk pre-filter (zero-hit chunks skip LLM extraction; safe bypass guards; alarm denominators adjusted) + shared per-chunk attempt budget (GRAPH_CHUNK_MAX_ATTEMPTS=2 across schema re-roll/fallback/429-replay; was 3-4 full passes ≈ ~24 provider calls per failing slot); run-level provider-outage breaker (LLM_OUTAGE_HALT_CONSECUTIVE=10 → existing failed→resume checkpoint state, no more 26h grinds); research attempt spend flushed to the meter exactly once on failure/timeout/cancel (fixes 31.9M vs 5.35M); _SLIM_EDGE_KEYS temporal fields; deployed skill overlay re-synced (byte-identity suite caught post-checkpoint drift — deer-flow/ is gitignored runtime, fixed on disk via runtime_skill_sync).
- **Tests:** full backend suite exit 0 / zero failures; 57 new red-first tests; 464 graph-adjacent + 318 orchestrator-suite green; ruff clean.
- **Next:** remaining backlog — research thread compaction/caching recommendation doc (product-behavior knobs need user sign-off), sim resume reuse, sim-subprocess meter persistence, chart theme unification, Step4 chart gallery. Expected impact so far: graph stage ~8h→bounded (filter + caps), outages halt in minutes not days, telemetry now complete for graph + failed research attempts.

### 2026-08-17T21:22:00Z — branch `codex/shenzhen`

- **Feature:** continuous improvement loop, iteration 5 (commits 0c2348f + 97fe0f5)
- **Landed:** sim resume reuse (three-layer fail-closed identity incl. LOOP-011 seal authority; SIM_RESUME_REUSE_COMPLETED=true), hollow-run completion gate, sim-subprocess meter persistence (last token blind spot closed; exactly-once orchestrator record); docs/RESEARCH_STAGE_OPTIMIZATION.md (owner-gated 78M-token levers); Step4/Step5 manifest-driven chart gallery + literal '![' strip; frontend dist rebuilt so :5001 serves the latest UI.
- **Tests:** full backend suite 3,003 collected exit 0 (sim agent's own gate); frontend 53/53; all focused suites green.
- **Loop total so far:** 7 commits (i1 UI entry/serving/failover, i2 telemetry/purity/polymarket, i3 dataviz, checkpoint, i4 graph bounding + outage halt, i5 sim fixes, i5b gallery). Every audit finding-cluster now addressed or owner-gated.
- **Next:** iteration 6 candidates (safe残余): research/report chart theme unification (forecast-visuals render.py now committed), self-hosted fonts, disk-usage report (no deletion — 3.2GB uploads retention needs owner sign-off), then loop wind-down pending owner review of RESEARCH_STAGE_OPTIMIZATION.md decisions.

### 2026-08-17T22:08:00Z — branch `codex/shenzhen`

- **Feature:** continuous improvement loop, iteration 6 (final)
- **Landed:** research-stage charts unified to the WAVE9 report theme (15 constants AST-verified identical; superior date-axis/dashed-forecast behaviors preserved; role palette re-validated; deployed overlay re-synced byte-identical); self-hosted fonts (@fontsource, CDN links removed, zero external refs in built page); backend/scripts/disk_usage_report.py (read-only; found 1.57GB of the 1.9GB reports footprint is 3 pre-i2 inline-plotly reports).
- **Loop wind-down:** safe-autonomous backlog exhausted. Owner-gated remainder: research token levers (docs/RESEARCH_STAGE_OPTIMIZATION.md), disk cleanup (report script ready), legacy-tree deletion, port-80 Apache (sudo), prompt-caching spike (paid provider experiments).

### 2026-08-18T05:55:00-05:00 (CST) — branch `main`

- **Feature:** continuous improvement loop, iteration 7 (loop re-opened by user /loop re-invocation; prior i1–i6 commits already merged to main)
- **Study→act:** the re-stated 80M-token and Polymarket asks re-opened the two matching owner-gated clusters. Three parallel agents (disjoint file ownership) + main-loop slices.
- **Landed:**
  - Polymarket correctness (agent P + main-loop closures): P0 influence boundary — FORECAST_MARKET_DIVERGENCE_MIN_CONFIDENCE=0.6 gates revision eligibility; irremovable market_influence stamp; reconciliation restores prior probability when the influencing anchor is judged wrong (partition-governed and superseded probabilities never rolled back); influences surfaced into market_comparison.json + bilingual report block. Reversed-token: name-based proofs pinned; producer rows persist outcomes/outcome_prices/clob_token_ids/clob_yes_token_id; deerflow_research price-history now charts the YES leg via _pm_yes_leg_token (fail-closed, never clob_ids[0]). Resolution-ledger check-then-append race reproduced (14 rows vs 9 correct) and fixed (lock+flock). "True price branded fabricated" incident closed: critique material now includes the market pack + explicit rule; final quantitative grounding gained market_supported status (snapshot-price boundary-exact match in market context ⇒ preserved, never an invented [S#]).
  - Research token levers (agent R): prompt caching was FORCE-DISABLED on the Claude OAuth path — re-enabled at the _get_request_payload choke point (≤4 breakpoints: system/tools anchor + newest-3 messages; thinking excluded; strip-first idempotent), default ON with DEERFLOW_CLAUDE_PROMPT_CACHE kill-switch (off = byte-identical legacy payload, pinned). Usage accounting verified cache-inclusive. RESEARCH_TRIM_TOKENS_TO_SUMMARIZE compaction knob wired, default preserved exactly. New tracked-vs-deployed byte-identity guard for patches/models+middlewares. Open: live billing verification of cached reads needs a paid spike (doc'd in RESEARCH_STAGE_OPTIMIZATION.md).
  - Frontend polish (agent F): 21-finding package — extended tokens (type scale/shadows/motion/state surfaces), prefers-reduced-motion, print stylesheets, unified empty/loading vocabulary, DossierViewer tab-wrap fix, GraphPanel chrome moved onto DRF tokens (chart encodings untouched), Step1–5 wrappers rebranded + bilingual, modal transitions + aria-labels. Entry bundle +0.14% JS / +2.3% CSS; zero external refs; dist rebuilt and verified served.
  - Main-loop integration: LOOP-017 defect 3 fixed (PipelineManager.resolve_handoff_dir with realpath containment; all 8 research.py sites + fork chart_* fallback — forks' dossier/translation/PDF/progress/charts now serve the shared base artifacts); LOOP-017 defect 2 fixed (/status now splices live heartbeat/owner/ETA/staleness/spend/budget via the previously caller-less helpers); start.sh readiness signature DeepAgentForecast→DeepResearchForecast (launcher exit-1 blocker) + test fixture; scripts/setup_port80.sh prepared (sudo pf 80→5001 loopback redirect + Apache disable, verified surgery on a pf.conf copy, --status/--uninstall) — awaiting user sudo run.
- **Tests:** full backend suite exit 0, zero failures (background gate log scratchpad/full_suite_i7.log); new red-first suites: market influence 18, market evidence 6, fork handoff 10, status live 4, prompt cache 13, trim env 8, yes-leg 3, ledger+markets+bridge additions 11; frontend 53/53 build green; ruff clean on all changed files.
- **Next:** i8 candidates — frontend display of the new status `live` block (spend/ETA/staleness in ResearchView); visual browser QA of F's polish (Chrome extension was disconnected); Polymarket resolution-monitor scheduling; first live-run verification of prompt-cache hit rates (owner-gated, paid); disk cleanup + legacy-tree deletion still owner-gated.

### 2026-08-18T06:12:00-05:00 (CST) — branch `main`

- **Feature:** continuous improvement loop, iteration 8
- **Landed:** (a) same-message-ID subagent token undercount fixed via a new fail-closed sub-overlay in apply_subagent_overlays.py — per-id positive-delta accounting replaces first-seen-wins (red-baseline test reproduces the historical 10→110-counts-10 drop; deployed tree patched; vendor-bytes overlay test green); (b) resolution monitor got its first-ever scheduler — RESOLUTION_MONITOR_AUTORUN_HOURS (default 0=off) subprocess loop started only from run.py, plus POST /api/research/resolution-monitor/run with in-flight dedupe; (c) run-vitals strip in ResearchView (agent F2): elapsed/ETA/liveness/spend/budget from the i7 live block, 20 new frontend tests (73 total), +2.1KB gzip, dist served; (d) terminal-elapsed semantics frozen at source (was age-since-now — 33-day elapsed on an old run), omit-on-missing timestamps.
- **Tests:** full backend suite exit 0 zero failures (scratchpad/full_suite_i8.log); patches 23, sync-guard 14, autorun 6, live-block 5, frontend 73/73; ruff clean.
- **Next:** i9 candidates — visual browser QA (needs user's Chrome), UI trigger for the resolution monitor + monitor_report surfacing, first live prompt-cache verification (owner-gated), binary-forecast first-class table in report UI, disk cleanup (owner-gated).

### 2026-08-18T06:28:00-05:00 (CST) — branch `main`

- **Feature:** continuous improvement loop, iteration 9 (final active iteration; safe-autonomous backlog exhausted after this)
- **Landed:** viz_manifest.json provenance block (canonical-JSON sha256 per input artifact + policy descriptor hash; additive, unserializable keys honestly listed); .env.example exhaustiveness restored (11 loop-era knobs documented; check_env_drift.py fully green); BinaryForecastTable in ForecastReport (agent F3 — sealed forecast payload, zero new requests, bilingual, per-cell degradation, market-influence badge surviving anchor ejection; 13 new tests, 73→86); resolution-monitor UI trigger in SettingsMenu Maintenance (202/409/error inline notes).
- **Tests:** full backend suite exit 0 zero failures (scratchpad/full_suite_i9.log); provenance 4 + manifest API 7; env-drift 18; frontend 86/86 build green +2,692 gzip bytes, dist served.
- **Loop state:** every LOOP-017 defect cluster now closed or owner-gated. Remaining items ALL need the owner: sudo port-80 script run, RESOLUTION_MONITOR_AUTORUN_HOURS enablement, paid prompt-cache verification run, disk cleanup (1.57GB identified), legacy-tree deletion, visual browser QA (Chrome extension), confidence-chip tone unification (preference).


### 2026-09-07T15:05:20+00:00 — ASTRA documentation audit

- Scope: User-requested ASTRA-RECOMMENDATIONS.md and the session's recurring-workflow review.
- Outcome: Current architecture map, historical accounting reconciliation, 20 prioritized recommendations, validation criteria, and a single narrow forensic skill.
- Evidence: docs/research/astra-evidence.json; focused continuity record docs/handoff/astra-recommendations/handoff.md.
- Validation: 138 local link occurrences checked, 56 cited files hashed, Markdown parsed, skill validator passed, independent frontend/operations document review and raw-artifact skill forward test completed. No application build/test run was needed or claimed.
- Files: ASTRA-RECOMMENDATIONS.md; evidence JSON; versioned skill source; focused handoff; this progress entry. Root ignored handoff received append-only notes.
- Remaining: ASTRA-01 through ASTRA-20 are recommendations, not implemented features. Next session should rebaseline and plan ASTRA-01/02 with evidence-preservation and summary-provenance tests before coding.
- Delivery: audit artifacts committed on codex/astra-workflow-review-2026-09-07 using an isolated worktree. Push to origin was rejected for invalid GitHub credentials. Runtime/source changes were not included; remote publication awaits authentication repair.

### 2026-09-28T06:15:00+08:00 (CST) — branch `feat/research-engine-v3`

- **Feature:** deep-research engine v3 — first-principles overhaul of `deerflow_bridge/linear_research.py` + new `deerflow_bridge/research_gateway.py`, bridge overhaul of `deerflow_research.py`, provider prompt caching, parent integration (feature id `feature_research_engine_v3`).
- **Design:** plan (scope/scout/plan) → per-KIQ bounded tool agents (step/novelty/context/budget/deadline stops, ≤3 tool calls per step) → budget-gated gap rounds → section writers + executive summary → deterministic QA → finalize. Cache-disciplined layout [ENGINE_CORE][RUN BRIEF][shared ctx][task], append-only agent conversations, prime() before every fan-out, GLM per-call extra_body, Claude ack turn so breakpoints land on the shared prefix, cache-aware `[usage] … cached=` metering. Durable resumable phases under `out_dir/v3`; exit 2 (resumable) only for real outages / no model output / no evidence.
- **Verification:** 3 adversarial review rounds (22 + 15 + 46 confirmed findings, all fixed or documented as deliberate partials); every fix has a regression test that fails on the pre-fix code; full backend suite green (see docs/handoff/research-engine-v3/handoff.md §8/§10 — the file is gitignored and stays local); offline E2E under the deer-flow venv rc 0; real parent-pipeline acceptance with a child process; GLM wire bodies unchanged; ruff clean; env drift strict rc 0.
- **Also fixed:** test hermeticity (.env no longer leaks into pytest), the backend Chinese probability parser ("概率 15%" read as 1.0), parent research lint never running on single-lane runs (now audit-only there), legacy extract-only salvage corrupting positional sources.json / v3 meta.
- **Next:** owner-gated live GLM run to confirm cache hits and A1/A2 (see handoff §9); follow-ups outside v3 scope: report_lint issues on manifest-owned multi-lane legacy runs, position-aware number verification.

### 2026-09-28T06:45:00+08:00 (CST) — branch `feat/research-engine-v3`

- **Feature:** `feature_research_engine_v3` residual fixes after the v3 commit.
- **Changes:** percent-aware number verification (a fact's "15%" needs a page percentage — written as one, a range bound, or a number in a percent-marked table/row/header block — not any bare 15); degradation events for template plans from an unusable/failed planner answer; finalize writes sources.json before research_report.md.
- **Verification:** backend/tests/test_research_engine_v3_residuals.py (fails on the committed engine); adversarial measurement on 4,775 real cached pages + 27 real reports: 0 real-layout regressions, 98.1% of coincidental percentage matches removed, 6 report flips all correct removals; regexes linear; full backend suite re-run (see handoff §10).
- **Next:** unchanged — owner-gated live GLM run; push after GitHub re-authentication.

### 2026-09-28T13:10:00+08:00 (CST) — branch `feat/research-engine-v3`

- **Feature:** `feature_research_engine_v3` — owner-approved live validation (now `passes: true`).
- **Live run:** quick depth, GLM-5.3, real DeerFlowResearchRunner → exit 0 in 543 s; 40 provider calls; 222k input / 78k output tokens; 78.4% of input served from provider cache; parameters accepted (no 400, no degrade); 4 KIQ agents stopped on their own (no fallbacks); QA 9/9; 28k-char report, 25 sources, 13 actors. (Last legacy run: ~20 h, 46.8M input tokens.)
- **Follow-ups applied:** fiscal-year ranges not verified as quantities; writers prefer fetched sources; unusable JSON replies quoted in their warning line.
- **Next:** push after GitHub re-authentication; optionally drop the obsolete RESEARCH_LINEAR_MODE=salvage from .env.

### 2026-09-28T14:40:00+08:00 (CST) — branch `feat/research-engine-v3`

- **Follow-ups:** unit-aware number verification in v3 (power/energy/currency; 54/425 real-report coincidences removed, 0 real regressions); report_lint never deletes, invents or truncates [S#] (pass brackets, citation-aware dedup, ranges, per-source guard); test fixes (fabricated page-number fixture, SIGINT handler independence).
- **Verification:** independent skeptics per change (their residual findings fixed and pinned by tests); new tests fail on the previous code; full backend suite (see handoff §10); ruff clean.

### 2026-09-28T18:05:00+08:00 (CST) — branch `feat/research-engine-v3`

- **Live run 2 (owner-approved, standard depth, GLM-5.3):** exit 0 in 30.5 min; 94 calls; 693k input / 123k output tokens; 81.4% of input from cache; 10 KIQs, 116 facts (69 VERIFIED), QA 9/9, 48.5k-char report, 34 sources (grounding 0.56 vs 0.36 on the quick run). Retries exercised live (timeout; reasoning-exhausted writer cap).
- **Tuning:** default GLM write cap 12k → 20k; search failures flag a degraded run only at ≥ 20% failed.

### 2026-09-28T18:40:00+08:00 (CST) — branch `feat/research-engine-v3`

- **Live run 3 (owner-approved, deep depth, GLM-5.3):** exit 0 in 39.3 min; 155 calls; 1.45M input / 244k output tokens; 81.3% of input from cache (~973 credits); 17 KIQs, 0 fallbacks; 194 facts (106 VERIFIED); QA 9/9 with 3 substantive critique rewrites; 92.8k-char report, 65 sources (grounding 0.63), research_quality 0.73. No code change needed.

### 2026-09-28T19:40:00+08:00 (CST) — branch `main`

- **Merged to main and pushed (3fc4f84):** research engine v3 (feat/research-engine-v3), removal of the Ling-Fin proposal (document deleted, paused code stash dropped), and PR #2 (claude/epic-keller-9ls0bz: README rewrite EN + 中文, dev-server auth gate, agent-log publication gate, per-provider research models, dossier edits surviving Continue, portable KG MCP config, demo-site fixes). GitHub marks PR #2 merged.
- **Merge:** 4 conflicts resolved to v3 (the PR patched the replaced v2 engine); its v2 tests retired with their intents ported to v3 tests; review follow-up: an OSError deploying the optional MCP config no longer skips syncing required bridge modules. READMEs updated for v3 in both languages.
- **Verification:** PR #2 reviewed by 4 area reviewers + skeptics and its own suite (no PR-caused defects); main: backend 4,255 passed / 0 failed / 11 skipped, frontend 72/72.
- **Follow-ups (pre-existing, not PR regressions):** unpublishable-report drafts still visible via outline.sections[].content on GET /api/report/<id>, /list, /by-simulation; section endpoints fail open when meta.json is unreadable; the agent-log in-progress exemption never expires for crashed reports; Vite dev server with host 'localhost' binds ::1 only on this Mac (http://127.0.0.1:3000 refused).
  → all fixed in 3d36ef4 (see the next entry).

### 2026-09-28T21:05:00+08:00 (CST) — branch `main`

- **Fixed and pushed (3d36ef4):** unaudited report drafts are withheld on every endpoint — outline section content on GET /api/report/<id>, /list, /by-simulation; one fail-closed gate for /sections, /sections-partial, /section/<n>, /agent-log(/stream) (missing, unreadable or non-object meta.json, or a publication check that raises); the in-progress exemption expires after 30 min without generator activity; /console-log(/stream) redact quoted draft excerpts while keeping summaries. Vite defaults to 127.0.0.1 (both http://127.0.0.1:3000 and http://localhost:3000 work, loopback only).
- **Verification:** independent skeptic audited every report route (found the console-log leak and the non-object meta.json 500, both fixed); new tests fail on the previous code; backend 4,283 passed / 0 failed, frontend 73/73.
- **Operational note:** a dev stack started at 17:48 from pre-merge code (vite --host on *:3000, backend without the forwarded-header gate) must be restarted on current main to close LAN/Tailscale access.

### 2026-09-29T20:00:00+08:00 (CST) — branch `main`

- **User report:** "Translate to 中文" on report_d79a064bf5cc (pipe_6c4190b31f0b) failed with "translation contains 1 target-language contamination lines"; the dashboard looked translated and the rest stayed English. Follow-up: make translation robust for other completed runs too.
- **Root causes:** (1) GLM renders "December 2024" as "2024年12月", adding a month numeral the English source lacks, so every numeric-integrity guard rejected those sentences (one stayed English → the fail-closed audit withheld everything; accepted fallbacks silently dropped months — report_54f0a34a90b6's published zh lost all 22 month dates). (2) Retry replayed the in-memory LLM cache byte-for-byte. (3) forecast.json spine/critique prompts had no output-language rule → English runs got Chinese headlines/scenarios (dashboard + scenario chart). (4) The report UI's verified-translation contract required `citations_path`, which the backend never sent → the language toggle never appeared. (5) HTML comment markers rendered as text. (6) Duplicate translate POSTs blocked on the generation lease; a backend restart stranded "generating" for 15 min; a dead HTTP/2 stream stalled translation 10 min (600 s read timeout).
- **Fixes:** `translation_dates.py` (dates → atomic target-language tokens; date-aware fact parity everywhere cross-language); numeral-word rule for 中→EN; cache bypass + temperature ramp; whole-line last-resort repair; bounded residual tolerance (`REPORT_TRANSLATION_RESIDUAL_LINES`, 0 = strict) with UI note; forecast language rule; `forecast.<lang>.json` + `/forecast?lang=`; dashboard follows the view language; zh chart captions/alt text; zh PDF labels (目录/图/表); `citations_path` in the verified row; comment markers hidden; duplicate POST answered before the lease; dead-owner → interrupted (report + dossier); translator version + `outdated` + `?force=1` "Update translation" (published variant kept on failure); provider-failure counts in rejection messages; 240 s per-call timeout for translation.
- **Verification (live, GLM-5.3):** all five publishable completed runs now translate — report_d79a064bf5cc en→zh, report_9147b3f6a0a9 / report_970e5aa53841 / report_c83f21765b96 en→zh (all three failed in July), report_ffe1ea6bf50d zh→en (21 sections, dashboard 114/114); every audit hard-passed with 0 residual lines. "Update translation" on report_54f0a34a90b6 (old engine had dropped all 22 month dates) regenerated it in ~4 min: 22/22 dates, no longer outdated. Headless-Chrome checks of both language views and of the update flow; zh PDF renders with 目录/图 labels. Tests: backend 4,356 passed / 0 failed / 11 xfailed, frontend 76/76, vite build OK.
- **Also fixed while verifying:** fused letter-digit identifiers (FY2030, Q3, H100) kept verbatim; integrity retries and dashboard localization run concurrently; a live owner is trusted through long sections; the update run no longer hides the published translation in the UI; dossier "interrupted" state offers a retry.
- **Follow-ups:** re-render chart images per language; 27 older completed reports are withheld by the publication gate (no final audit / stale policy) — a deterministic re-audit tool is needed before they can be translated; report_ffe1ea6bf50d's primary PDF fails the numeric-token integrity check; push blocked on GitHub re-auth.

### 2026-09-30T01:35:00+08:00 (CST) — branch `main`

- **User request (2026-09-29 22:38):** update the live demo site with the most recent run in Chinese and English; commit, push and merge to main (GitHub token supplied in chat; used only for the push commands, never stored).
- **Published:** quantum-2040 (pipe_6c4190b31f0b, GLM-5.3) with English + Chinese report and research dossier; grid-storage-2040 gains its Chinese report; datacenter-2030 gains its English report (and its three charts back). Report/dossier tabs switch English/中文 and follow the site language.
- **Engine fixes found while producing the translations (b26c819):** (1) magnitudes — with numerals frozen the model wrote "$3.7 billion" as "3.7亿美元" (10× too small) and "$255 million" as "255万美元"; the published zh quantum/grid reports carried it. Amounts are now atomic tokens rendered in target units and checked by value (translation_quantities.py; translator 2026-09-29.2). (2) Every dossier translation failed: lint-before-audit moved References after the Visual Annex, and zh lint deleted sentences containing 轮次 (融资轮次, funding rounds) as simulation leakage. (3) The dossier path now shares the report path's concurrent translate-and-retry engine; rejected translations keep their text for diagnosis.
- **Exporter (fe5befd):** language versions + maps in meta.json, `--reports-only`, `--graph-api`; assets resolved before a folder is replaced; a refresh keeps a published chart only while it matches the manifest.
- **Site (5f8585a):** EN/中文 switch, quantum card, 15-run READMEs, long reference URLs wrap on phones (pages scrolled sideways at 390 px before).
- **Live translations (GLM-5.3, new engine):** quantum report en→zh and dossier en→zh, grid-storage en→zh (amounts 38/38 and 18/18 value-equal to the source; RMB 250 billion now 2500亿元), datacenter zh→en (third attempt: attempt 1 exposed the year-range misread "in 2024 to $200 billion", attempt 2 lost one H2 in post-processing, cause not found — rejected text is now kept for the next occurrence; whole-document facts identical, 101 amounts). All audits hard_passed with 0 residual lines.
- **Verification:** backend 4,406 passed / 0 failed / 11 xfailed; headless Chrome on a local copy of docs/: both languages of report and dossier, all images load, site-language toggle, single-language runs unchanged, no console errors, no sideways scroll at 390 px.
- **Follow-ups:** residual-line repair partially translates English reference titles inside Chinese variants (References should be exempt); chart images keep English labels (and quantum-2040's scenario chart shows the Chinese scenario names its forecast.json was sealed with); datacenter-2030's dossier contains a "[S42](K6 口径…)" pseudo-link that the strict exporter rejects, so its dossier is not re-exportable as-is; the H2 lost in datacenter attempt 2 is undiagnosed.