Improve Alden graph reading, collection and model routing

[twoimo]

Alden keeps graph exploration readable with a navy background, useful zoom-dependent labels, bounded rendering and reversible source reading. Numeric facts remain available, display duplicates do not consume useful-fact slots, and source identity, relation provenance, project scope and model routing settings remain preserved.

Alden 0.3.154 fixes a reproduced refresh failure: early Threads inputs without an observation timestamp caused the installed receiver to reject a genuinely newer snapshot. The compatibility path accepts a legacy input only when its retained capture, exact bytes, current membership and collector read receipt all agree. Incoming observation time remains mandatory; missing or corrupt proof, explicit null timestamps and stale input still fail closed. It never invents an old site observation time.

The installed app completed one existing Threads target through collection, capture and E5 indexing: added 0/revised 1, two text windows, the same document ID and a new retained version. Installed stdio MCP search/graph/history returned that version; search used warm E5 RRF. Replaying the completed request retained the original request/attempt and did not requeue. All 64,395 previous versions, 118,555 relation-history rows and 48,355 relations matched the consistent pre-run backup, as did non-target documents and memberships.

The Graphify v0.9.92 audit pins official revision 86c57be and compares selective code/provenance adapters against a full engine replacement. Isolated tests reproduced generic hub labels, four identity/quantity contract conflicts in six hand-authored cases, unrelated-component partition sensitivity on Louvain, and stale extraction after a same-size edit restoring an old mtime. AST and explicit edge-export controls passed. These are scoped counterexamples, not population error rates. The full 63,019-byte GPT-6 Pro Deep Research thesis and original UI/hash receipts are retained; findings map to decisions, evidence, conditions and verification. The new Gemini adversarial review is queued without upload/send, and final reciprocal validation remains incomplete. Graphify has not replaced the product engine.

Validation: Python 165 focused collection/acquisition methods, UI 983, Rust 146 and strict Clippy passed for runtime source 659d3b5. All four source CI jobs 39161502783 passed. The 77 installed files/modes match the candidate; stable strict/deep signature, independent 0.3.153 recovery, protected settings and seven target policies reconcile. Candidate/installed/post-collection owned WKWebView audits passed with 34 captures per run; selected screenshots were directly inspected. The canonical build receipt remains valid. Prior Graphify documentation CI 39159342140 also passed all four jobs; the final documentation commit d74c08a also passed all four CI jobs 39163039860.

The installed retrieval path was additionally evaluated on all 213 Korean MIRACL dev queries and 3,057 human relevance judgments over the union of 2,835 judged passages. Actual E5 encoding used 720 bounded local requests for 5,755 inputs; all 2,835 producer float32 vectors matched pooled real responses. BM25/Dense/raw RRF/product consumer nDCG@10 was .5252/.7213/.6828/.6802. Official NIST trec_eval v10.0 matched all 5,992 per-query and aggregate metric values. This is a judged-passage subset baseline, not full-corpus or private-source quality certification; unjudged exposure is reported separately. No product weighting/model change was made. A final isolated-store cleanup failure was repaired from retained query receipts with zero query replay or new encoding. Reproducible export, frozen-input evaluation, eight CLI rejection controls and protected installed-state readback are documented.

Public retrieval audit commit 215992c passed all four CI jobs 39165781030. The canonical build receipt remains valid for installed runtime 659d3b5/0.3.154; the audit did not require rebuilding or reinstalling the app.

Frozen retrieval stage tracing reproduces every original ranking for all four methods across all 213 queries. Five added stages isolate product lexical candidates, ID/collection-time ties, body evidence and final admission; NIST checks all 3,210 additional metric values. Body preference improves nDCG by .0013343 on this restricted regression, and post-body top10 equals the consumer for every query. No new embeddings or network attempts occur during the replay; the entire original audit directory remains byte-identical. Four CLI rejection controls pass. The tool and decision matrix are in commit26f8b99.

A separate 44,791-byte research report is preserved verbatim with URL, source/bundle hashes and 27 original UI text receipts. Its self-identification as GPT-5.5 Thinking conflicts with the observed GPT-6 Pro/Deep Research UI. Backend identity remains unverified and this report is not certified as fulfilling the strict requested model tier. Thirteen findings are independently mapped to source, executable checks or explicit conditions. Original accepted usage and completed-slot release remain recorded; no retry, model substitution or final cross-audit completion is claimed.

All 18 goal requirements and Cursor's eight original cross-gates remain active. Independent human semantic/identity/time gold, physical UI/voice/Retina, two globally connected knowledge-MCP hosts, broad acquisition, whole-app resource measurements, valid worker activation proof, completed external reciprocal audit and notarized production release are still open. No source merge, model substitution, live-send test or production release is claimed.