v2.6.0 — Technical context across AI research

v2.6.0 — Technical context across AI research

Published on 2026-11-05 from the approved merge of release PR #256 (fccb6b3). All three release-preparation CI jobs passed: build-and-test, clean-install and postgres-contract. The changelog heading records preparation on 2026-11-04; publication is 2026-11-05.

The same three descriptive observations now support AI Investment Briefing, Interrogate, Second Brain and selected holdings in Portfolio AI Risk Assessment. Exact values, dates, formulas, source identities and missing coverage remain inspectable. These observations do not establish historical investor knowledge, a long-term buying case or aggregate portfolio risk.

Upgrade

Includes Alembic 0018_brain_technical_context: one nullable JSON column for explicitly saved Brain technical snapshots. Docker applies migrations before startup. Existing local databases must run PYTHONPATH=. backend/.venv/bin/python -m alembic -c backend/alembic.ini upgrade head with their configured database before using the new app. Old memos are not backfilled or rewritten; downgrade retains captured research. No new provider, indicator, dependency or embedding change is required.

Verification

Consolidated checks passed 1,667 backend tests (one skipped; 77.29% coverage), all 681 frontend tests, the production build, API/surface/mock guards and all 24 Chrome smoke journeys. Compose configuration and disposable SQLite legacy-memo upgrade/rollback/re-upgrade checks passed. PostgreSQL and fresh-stack contracts passed in CI. The maintainer confirmed the Brain correction and Portfolio integration work as intended; automated fixtures are not an independent model-quality benchmark.

Fixed

• Reliable manual Brain symbol selection — Add stays aligned with the input
  when suggestion help disappears on blur, preventing a lost click and an
  unintended ask without technical symbols. Browser coverage waits for empty
  suggestions and verifies each selected symbol before submission.
• Brain technical scope survives missing private matches — explicitly
  selected technical symbols are collected even when private retrieval has no
  matches, the index is empty or query embeddings fail. The answer reports the
  private evidence gap, keeps available technical citations inspectable and
  preserves captured evidence on save. Private citations are required only when
  supplied; invented references still fail. Prompts keep long-term investment
  conclusions beyond the scope of three short-window measures. If private and
  technical evidence are both unavailable, no synthesis is requested.
• Browser insight length parity — the shared insight reader now accepts the
  server's expanded summary/title/point and total-text safety ceilings, instead
  of rejecting valid longer analysis with a generic request error. Unicode
  counts match Python code points; insight-specific NDJSON framing accommodates
  escaped Unicode without enlarging other model-operation frame limits. A
  backend/frontend contract check prevents the limits drifting apart again.
• Actionable structured insight repair — validation failures now identify
  the field and safe type/length metadata without logging provider text or
  private evidence. Streaming fallback and the existing single completion retry
  receive the validation correction instead of repeating the same prompt.
  The inherited 220-character bullet limit is now a writing preference rather
  than a rejection rule. Generous safety ceilings allow 2,000-character summaries
  and points, 160-character titles, 16,000 total text characters and 64,000 raw
  JSON characters; oversized streams close before bounded fallback. Structure/
  count checks, selected model, shared deadline and final validation remain;
  malformed object/list tone values degrade rather than crash. Text is never
  silently truncated.

Added

• Selected holding technical evidence for Portfolio AI Risk Assessment (v2.6)
  — optionally choose up to three held symbols alongside existing risk metrics.
  The server verifies owner/membership before market or model work, aggregates
  recorded lot quantities and cost-currency metadata, and collects a cancellable
  15s batch using the same three observations. Per-holding references inspect
  exact dates, formulas, provenance and supporting closes, including partial
  coverage and model failures. Coverage counts symbols; no value weights or
  aggregate technical risk are computed. Missing risk metrics do not block
  explicitly selected technical evidence. Scope/holdings/metric changes cancel
  and clear old analysis, and late analytics from a prior portfolio are ignored.
  Existing risk APIs add typed optional scope/evidence; supplied risk observations
  are bounded at 32,000 serialized characters and preserved without truncation.
  Deep/default profile, deadlines and bounded repair remain. No new indicator,
  provider, route, persistence, indexing or migration is introduced.
• Optional technical evidence for Second Brain (v2.6) — choose up to three
  explicit instrument symbols alongside existing private-source filters. One
  bounded batch supplies the same three descriptive observations, alongside
  private research when available. Namespaced metric citations inspect exact dates,
  values, formulas, provenance and supporting closes separately from private
  citations. Current snapshots stay distinct from dated claims and historical
  knowledge; unavailable coverage preserves private evidence. Same-profile
  streaming/completion reference repair stays inside existing deadlines, and
  cancellation joins provider work. Explicit Brain Log saves retain the captured
  evidence; private citation rechecks never refresh it. Reviewed Note promotion
  includes dated capture labels rather than reusing active snapshot references.
  Nullable JSON migration 0018_brain_technical_context leaves old memos intact
  without backfill/default and retains the column on rollback. No new route,
  provider, indicator, retrieval-ranking change or automatic indexing is added.
• Technical context for instrument research (v2.6) — Investment Briefing and
  Interrogate gather the three standalone observations during explicit
  generation. Answer-local metric references open exact dates, formulas,
  provenance, quality limitations and supporting closes. Missing technical
  history preserves other research inputs; reference checks share the existing
  bounded Deep repair/deadline lifecycle. Cache keys include evidence contract/
  UTC date, and Interrogate retains owner/note scope and distinct effective/
  recorded note dates. Existing research GET/stream results add typed
  technical_context; no new route, indexing or persistence is introduced.
• Standalone technical snapshot foundation (v2.6) — internal typed
  instrument requests/snapshots select a bounded recent daily window without
  creating a comparison pair. Shared original-source checks and deterministic
  20/21-close calculations preserve provenance, actual dates, stable identities
  and unavailable states. Collection makes one bounded provenance-aware history
  call and propagates cancellation; no AI consumer, new public route, cache or
  persistence is introduced. Existing v2.5 comparison contracts remain intact.