docs(roadmap): roll forward to 2026-09-26 — the stale-evening corrections + rc.78 queue

[xerj-org]

Motivation

The pinned roadmap's last review (2026-09-21) was stale within hours:
#950 closed 19:33, #941 19:38, #1015 20:49 (all the same evening), #874 the
next morning, and the stage-1 gating trio #937/#938/#939 closed that evening
too — every one of them still recorded in ROADMAP.md as open, "in flight" or
"under way". This pass re-checks every status claim against live tracker
state and rolls the file forward. (Found by the 2026-09-26 post-merge audit.)

What changed

• Next release: the two "in flight" items (Flush publication guard spans the whole finalize — capture-retry latency under flush load (follow-up to #1013 / #1014) #1015, engine: RSS pinned above the memory watermark after bulk ingest into 1,526 indices — 14.8 GB anonymous memory for 1.2 GB on disk, not released at idle; node refuses every write until restart #950) are closed with
  fixes on main riding to rc.78. The section now records what actually
  landed since the rc.77 tag (fix(api): POST /{index}/_cache/clear releases the index's rebuildable caches (#950) #1009, fix(#950): id-position maps from the __id projection; stream the stored reassembly #1017, fix(#1015): flush drains freeze; the publication bracket wraps only the publish #1018, bench(#874): idle-budget fixture + CI gate; delete the O(N) metrics gauge loop #1020, fix(#1019): _delete_by_query pages the whole match set, ids-only #1021, fix(#1022): _update_by_query pages the match set; exact sig-text/enrich counts #1023, perf(#1024): lazy request-cache seen-set — 206 to 64 kB/idx idle RSS #1025,
  bench(systemone): measure local /_decide for discussion #1012 email labelling #1026, docs(llms): #1028 — status claims caught up with main (rerank/share/mbox on main; #948 closed; residual retention re-anchored to #1032) #1033, perf(#1024): request-cache seen-set allocates on first tracked search — idle 206 → 64 kB/idx #1034) with PR links — including Meta: idle cost must be near-zero on dev laptops — the reference-coding resource budget (#871 #872 #873) #874's budget met at ~3×
  margin and discussion Optional Jev classification in `xerj autoindex` (emails first): would you use it? #1012's email-labelling measurement.
• Open defects carried into rc.78: now autoindex: a source file that grows mid-run (append-only log) kills the run, is misclassified as a bulk/backend failure, and the retry advice points at the wrong fix #1031 and engine: stored_value_cache / dv_cache / id_pos_cache remain unbounded until merge-retire — the residual heap-per-doc retention after #1002 #1032 only; trackers
  sentence corrected (Tracking: the zero-token direction, stage 1 — rerank stage, share links + reading room, mbox/Takeout ingest, autoindex resilience #941/Meta: idle cost must be near-zero on dev laptops — the reference-coding resource budget (#871 #872 #873) #874 closed; only 📍 XERJ public roadmap — where the project is going and how to follow it #298 remains by design).
• GA gate: "Close the CHANGELOG gap" marked closed 2026-09-26 (backfill is
  PR docs(changelog): reconstruct the rc.19–rc.70 gap (GA gate item) #1035 — merge docs(changelog): reconstruct the rc.19–rc.70 gap (GA gate item) #1035 before this PR so the claim is true when this
  lands); the record-complete note replaces the gap warning.
• Zero-token direction: gating trio recorded as fixed (fix(api): fields/docvalue_fields pierce an explicit _source includes filter (#932) #991/fix(ingest): bounded window concurrency for neural _bulk embedding (#938 slice 1) #995/perf(knn): exact scan ranks addresses and hydrates only the winners (#939) #979) with
  old costs kept as dated before-states; hybrid RRF orders tied fused scores by HashMap iteration: top-10 differs for 21 of 40 queries after a restart on unchanged data #940 spread annotated pre-fix; stage-2
  object storage marked done (Wire the index segment path to the object-storage backend (bundle a segment into one object first) #965 wired rc.77 — the old bullet contradicted
  the file's own Shipping today); mail ingest "unreleased" → "shipped in
  rc.75"; three stale *In flight:* labels on landed items → *History:*.
• Review line → 2026-09-26, honestly scoped as a desk review; the
  milestones claim is true again (rc.78 milestone created;
  corpus/index: portable one-file bundles — indexed data that works across machines #1030/autoindex: a source file that grows mid-run (append-only log) kills the run, is misclassified as a bulk/backend failure, and the retry advice points at the wrong fix #1031/engine: stored_value_cache / dv_cache / id_pos_cache remain unbounded until merge-retire — the residual heap-per-doc retention after #1002 #1032 triaged to it, 📍 XERJ public roadmap — where the project is going and how to follow it #298 → GA — done on GitHub already).

Verification

Every closure/fix claim above was verified against the live tracker
(gh api issues/N state+closed_at) and PR merge records; numbers quoted only
from PR bodies' own verified runs (#1002, #1026, #1025/#1034).