Merge pull request #1016 from xerj-org/release/rc.77

release: cut 1.0.0-rc.77

release: cut 1.0.0-rc.76

The JEV-interface and honest-fixes release. The node now speaks the Jev
reranker's wire protocol locally; everything else is a fix with a
reproduction behind it:

- /v1/systemone + /_decide: a rank-weighted kNN vote over a
  labelled-history index, nothing leaves the node; accepted by the
  unmodified pip-installed jev-reranker under its own strict validation
  (#975)
- the rerank stage's shared-state bug fixed and measured: 0.3822 ->
  0.8299 nDCG@10 through the real stage path; SciFact 0.6572 -> 0.7410,
  NFCorpus 0.3016 -> 0.3312 across 3 shuffled repeats; FiQA BM25 0.2382 ->
  Jev 0.3638; AG News 0.9182 accuracy, no model, no tokens
- WAND fast path in xerj-fts: 2.0x on the vote, bit-identical (0 order and
  0 score-bit changes over a 648-query A/B; ES-YAML 0 failed)
- the reference-coding loop is in the binary: xerj code, xerj corpus
  add|index|list, the xerj_code_search MCP tool — no python3/curl left; the
  --fresh swap contract re-pinned as ten Rust tests (#977, #1004)
- ingest memory: a 1 GB mailbox no longer aborts under the 16 GiB cap,
  at-rest settled 972.8 -> 481-488 MiB (variant A) (#948)
- scalar8 written at ingest: 2.3x kNN, filter drift exactly 0.0 (#392)
- the Windows build compiles again (MSVC purge-hook cfg gate)
- .deb release assets, amd64/arm64, packed from the staged tar.gz binary
  (#810); Debian archive inclusion measured, not claimed
- fourteen more fixes with fail-before proofs: #962 #960 #935 #943 #959
  #933 #961 #932 #937 #938 #971 #939 #974 #954

Version stamps: engine/Cargo.toml + Cargo.lock at 1.0.0-rc.76; landing
footers V1.0.0-RC.76 (49 files) and the hero data-latest-tag seed.

Verified on this branch: cargo check -p xerj-server clean, xerj-autoindex
lib 1066 passed / 0 failed, SEO gates green (build_articles --check,
gen_sitemap --check 177 URLs, test_article_schema 5/5).

v1.0.0-rc.75 — object storage and freshness

Index an S3-compatible bucket (xerj autoindex s3://bucket/prefix), keep a
folder or a bucket current with --watch, share one index read-only with one
person, ingest mbox and Google Takeout, read any record in the console with a
guest mode, and reorder search results with an opt-in rerank stage.

A real S3 StorageBackend with a read-through cache ships underneath — the
client only. An index still cannot live in a bucket: storage.backend = "s3"
refuses to start, on purpose.

Every object-store path counts billed attempts rather than successes, keeps a
spend ledger that survives a restart, and stops at a budget that is
free-tier-safe by default.

Measured and deliberately not shipped: a local cross-encoder judge does not
beat the hybrid query XERJ already ships.

release: cut 1.0.0-rc.74

The adoption release. rc.73 made retrieval definition-first; this one makes
adopting it one command. xerj init detects the coding agents in a project
and writes the minimal wiring for each (MCP registration, a ten-line skill,
an AGENTS.md section) with backups and a dry-run; xerj gain reports what the
node actually did — counted from the hash-chained audit log, never
estimated. Full contents: CHANGELOG 1.0.0-rc.74.

release: cut 1.0.0-rc.73

The definition-first release. A cross-file retrieval study measured the
search CLI's plain multi_match finding the correct source file at rank 1
only 38% of the time while a defs-first bool measured 94% on the same 80
queries, so that shape is now the default everywhere a plain string comes
in: xerj search, and the MCP xerj_search tool, which previously rejected
plain strings with an HTTP 400 and forced agents to hand-write query DSL.
xerj def joins as go-to-definition: file:line + signature in one call.
Results are passage-first — never whole file bodies, ~9x fewer tokens over
three code hits. Full contents: CHANGELOG 1.0.0-rc.73.

xerj v1.0.0-rc.72 — the idle-cost release

Three of four idle-cost mechanisms fixed (#871 event-driven merge scheduling,
multi-index fan-out (#875), and six correctness fixes (#825 #830 #790 #781

Cut after a fresh open-PR sweep (0 open; all 11 PRs reviewed before merge,
5 blocked on real defects and repaired). Conscious exclusions: none.

Known regression, documented: the ES-native {query, knn} shape is correct but
slower than the hybrid query type; #892 carries the indexed-route fix. The

xerj v1.0.0-rc.71 — ES-compat correctness release

30 query/aggs/API semantics fixes, ~100x incremental re-index (#868),
metrics counters wired (#804/#819), runtime flood-stage unblock (#823).
Cut after a fresh open-PR sweep (0 open; all 12 outstanding PRs reviewed
and merged today). Conscious exclusions: none.

v1.0.0-rc.69

autoindex: actionable error on a legacy text-mapped catalog (#755/#756)
autoindex: per-file record cap + truncation reporting (#381/#758)