v4.15.0

Fixed

• The transcript panel published a net drawn from a 9-turn control. It
  printed 1,905 tok/turn cost in bold against an "ungrounded" cohort of 3
  sessions totalling 9 turns (1, 2 and 6), versus 100 grounded sessions with a
  median of 15. The cohorts are observational and confounded by session
  length — a session that never grounds is usually one that barely started, and
  the 9 sessions in NEITHER cohort, equally short, spent MORE per turn than the
  grounded ones. The sign was not evidence in either direction. _proxy_panel
  already withholds below its floor in this same module, calling a too-thin
  sample "not a measurement with a caveat — not a measurement"; that standard
  now applies to both panels instead of one getting the word "provisional".
• Duplicate title and repeated bullets in the dream profile. The distill
  prompt says "no top-level title" and asks for a distillation, but nothing
  checked either. Live _profile/profile.md: # Profile — global on lines 8
  AND 10, and 13 of 62 bullets literal duplicates — one line repeated 12
  times — in a document the briefing injects verbatim on every session, where
  it was 43% of the whole payload. render_profile now drops an echoed
  heading and collapses repeated bullets, order-preserving and deterministic
  (non-bullet lines are never deduped). Measured across the seven live profile
  documents: -9.2% overall, -23.2% on the global one, ~238 tokens per session
  for a briefing loading global plus project.
• The profile's identity section never said what the project was. The
  distill prompt lists identity as a heading without defining it, so the
  model filled it with whatever the recent sessions were about — a session
  still had to be told what the project is, which is the thing the profile
  exists to carry. Verified against the live store: a decision memory
  titled "Qué es memo (identidad del proyecto)" sat at row 1 of the 40-row
  source window and the rewrite still opened with last week's token work.
  identity now leads with what the project or system IS whenever a memory
  states it. After the change the memo profile opens with "a local-first
  semantic memory for AI agents. Stack: embeddings MLX + hybrid sqlite-vec…",
  and the SessionStart briefing carries it.
• ### Memory relations named nothing. It rendered - \492e8d82` related
  `715ed835`— two hex prefixes and a verb, the only graph output in the briefing a human could not read. Titles now come from oneget_batch`, and a
  row whose memories no longer resolve is dropped rather than shown as bare
  ids.

Added

• memo tokens reports a prefix counterfactual. The holdout A/B is the
  better instrument but its control arm fills per session at
  MEMO_PROXY_HOLDOUT_FRAC, so "not enough data to compare arms" was reading
  as "memo saves nothing" — a different and false claim. The transforms are
  pure functions of the payload: the volume removed is recorded per request,
  and the surviving prefix's own counters say what tier that volume would have
  billed at. Weighting the removed volume at that same observed weight needs no
  control arm. Measured on 10,398 live requests: 54.1%, 1.62B tokens of
  prefix removed at the 0.131x weight the surviving prefix actually billed at.
  Labelled "not an A/B" wherever it appears, and never substituted for one.