[wfaithfull]

This is step 1 of the 0.7.0 release (CONTRIBUTING.md, "Releases and upgrades"). It is the human-written text that the release commits leave unchanged.

CHANGELOG.md: the 0.7.0 section, checked against every PR merged since 0.6.0. The heading line is left for release/prepare.py to date.

• Added the missing entries:
  • external ids made row by row in walk/neighbours (Build rows lazily: walk and neighbours (M11) #66);
  • packages (Release automation: prepare workflow, gated publish, release scripts (M11) #65, Distribution: signed apt repository, Docker images, Homebrew tap, arm64 and macOS CI (M11) #67);
  • the cursor SIGSEGV fix (Leave a cursor an ERROR unwinds through to the abort (fixes a SIGSEGV under parallel query) #86);
  • oriented triangles (triangles orients edges by degree and finds each triangle once #87);
  • the contract counted-once sweep (Contraction counts a pair once in one sweep (fixes 0.7.0's Louvain/Leiden slowdown on graphs with relations) #85).
• "Upgrading" is reordered to match 0.6.0's style.
• ### Added is renamed ### New, as in earlier sections.

README.md Status:

• 0.7.0 is described as released, with its misses and caveats.
• 0.8.0 is listed as in development (M12).
• "Not there yet" now covers M12 only.

DESIGN.md §13: the owner's decisions of 2026-10-26.

• 0.7.0 is cut now. The corpus's large tier, the multi-day pooler soak, and Appendix A.1 and A.3 at their full edge counts move to M12, before the 1.0.0 release candidates.
• M11 now has its "Done (0.7.0)" list.
• M11's performance leads come after 0.7.0:
  • Leiden over the previous level's graph;
  • PageRank as a kernel;
  • row-return overhead;
  • text messages in gather;
  • the memory estimate for cross-graph distance('edit');
  • the dense bridge's eigen.

Not in this PR: the apt key fingerprint on README's <!-- apt key --> line. The release publish job refuses to sign without it, so it is needed before the first packaged release.

sql/schema-drift.sh and pytest release/tests pass.

---

Generated by Claude Code

[wfaithfull]

"Generators + SQL harness (PG 16)" failed in the pooler soak. In the steady window the cluster's progress_rows rose from 5 by about 3; the soak allows 2. Every pool median and every other cluster measure was flat or falling.

This doesn't come from this PR, which changes only CHANGELOG.md, README.md and DESIGN.md. It is the second soak failure today on a different measure: #89's first run failed on one backend's anonymous memory and passed on re-run. Main is green on 18295a5.

I'm re-running the failed job once. I'm also looking at whether the soak's steady-state bounds are too tight for the chaos clients it runs (cancels, and clients that abandon a call part-way), or whether a progress frame really outlives its call. Any fix will be a separate PR.

---

Generated by Claude Code

[wfaithfull]

"Generators + SQL harness (PG 16)" failed again, this time in crash recovery, not the pooler soak. The soak passed with #91's fix.

The failing step is test/crash_recovery.py (preload), "a refresh after terminate":

• It took about 2 minutes, where each step usually takes about 4 s.
• The test then reports the refresh after recovery built (nodes, edges) [['200000', '200000']], the source has [].
• Every other step passed, and the whole run without preload passed.

This PR changes only CHANGELOG.md, README.md and DESIGN.md, so the failure isn't its own. Crash recovery passed on this PR's previous run (6f61f91) and on main at 18295a5 and 6771e70. I'm investigating whether it is a real fault, a refresh after a terminated refresh in a preloaded server, or a race in the test. I'll re-run the failed job once when the run finishes.

---

Generated by Claude Code