The GitHub connector, live, for the first time.
Two known gaps fixed. A third found mid-run.

"Migration for #3828" — no keyword, no edge, before today.

PR #3834 body: "Migration for #3828"              // no keyword
PR #6606 body: "PR #6514 changed signup tracking..." // mid-sentence
PR #6597 body: "Extracted from .../pull/6591"       // full URL

github_analyzer.rs only detected `#N` immediately after a documented GitHub auto-close keyword (`closes`, `fixes`, `resolves`) — real PR bodies in this repo overwhelmingly don't use that vocabulary. Fixed: `find_bare_issue_numbers` now catches bare `#N` too. Separately, `GitHubApiClient` issued exactly one unparameterized request — GitHub's default (30 items, newest-first) would have silently truncated a repo with ~4,600 real issues/PRs to its 30 most recent. Fixed with standard `Link`-header pagination, bounded by an explicit page cap.

1,600 real issues/PRs. ~23 minutes. No mock in sight.

$ export EKOS_GITHUB_TOKEN=$(gh auth token)
$ export EKOS_GITHUB_OWNER=plausible EKOS_GITHUB_REPO=analytics
$ export EKOS_GITHUB_PER_PAGE=100 EKOS_GITHUB_MAX_PAGES=16
$ time ekos build
Build complete. Files observed (new): 2
Total objects in ledger: 2404
real  22m50.039s

$ ekos recover
github-analyzer complete  items=1600  edges=13580
GitHub issues/PRs analysed: 1600

The 23-minute wall clock is almost entirely the per-PR `list_files` call — ~0.5s each, sequential, no concurrency — a real, accepted cost for a one-time documented run, not engineered around. 13,580 `References` edges from 1,600 items: file-change links, closing links, and the newly-added bare-mention links, all real.

1,533 of 1,600 real items. One identity. Confidence 1.00.

35. 'plausible/analytics#5869: Reduce noise in 2FA...' (PullRequest)
    — 1533 objects merged, confidence 1.00
      • #6527: Bump fast-uri from 3.0.6 to 3.1.4 in /assets
      • #5469: remove salts logs
      • #6534: Publish tracker package 0.4.6
      • #5158: time-on-page: `imported_pages` new columns
      ...1,529 more, almost every real item fetched

Not a subtle edge case — the single largest fact about the compiled result. `Custom("Issue")`/`Custom("PullRequest")` objects are named `"{owner}/{repo}#{number}: {title}"` — every item in one repo shares the `"{owner}/{repo}#"` prefix, and unlike RFC 0060's earlier `Table` fix (a namespace only), this one also swallows the item's own number, making the shared, uninformative part proportionally dominant. Jaro-Winkler's prefix bonus inflated every pairwise comparison regardless of how different the real titles were.

1,533 → 0. The catastrophic case is gone.

// RFC 0060's extension point, given a third case:
if matches!(&obj.kind, ObjectKind::Custom(k)
    if k == "Issue" || k == "PullRequest")
    && let Some(hash_pos) = obj.name.find('#')
    && let Some(sep_pos) = obj.name[hash_pos..].find(": ")
{
    return &obj.name[hash_pos + sep_pos + 2..];  // title only
}

Same architecture already built for `Table`'s schema-qualifier strip — just a third case, not a new mechanism. Verified: the single 1,533-object group is gone from the re-resolved output. Not a complete fix: 1,439 real merge groups remain (largest 174) — dependency-bump PRs sharing a `"Bump X from Y to Z"` template and feature-prefix conventions like `"time-on-page:"` still cluster. Same reasoning RFC 0060 already gave for not chasing a complete fix: a real design decision with estate-wide blast radius, not a scoped bugfix.

#5158 was chosen as the demo before the run. It didn't survive.

$ ekos query object <canonical id for #5100>
Object: plausible/analytics#5100: time-on-page:
        Ingestion logic for engagement_time (PullRequest)
  Properties:
    excerpt: "The previous PR was merged into a
              dependant branch not master :facepalm:"
    // #5158's own body and migration-file evidence: gone

Even after the fix, PR #5158 ("time-on-page: `imported_pages` new columns" — this RFC's originally-chosen example) is merged into a real 16-object group of sibling "time-on-page:" PRs. The surviving identity shows only its own data — exactly RFC 0060's "surviving object swallows the rest" pattern, now confirmed on GitHub `PullRequest` objects too. Reported plainly rather than quietly swapped for a cleaner-looking example.

A real PR. A real file. One hop, live.

$ ekos query neighbourhood <PR #6421 id> --depth 1
Neighbourhood: 2 objects, 1 relationships

  plausible/analytics#6421: Assert how we store
  data imported from GA4 (PullRequest) [root]
  test/plausible/imported/google_analytics4_test.exs (File)

  References  #6421 → google_analytics4_test.exs

PR #6421 survived standalone — its real `References` edge to the real test file it changed is graph-traversable via `ekos query neighbourhood`, unaffected by the residual over-merge issue. This is the live, working version of "beyond Git/code": a real GitHub item connected to real code, not just co-existing in the same ledger.

Docs. Code. A GitHub PR. One search, no hand-curation.

$ ekos query find "google analytics import"
31 result(s):
  google-analytics-import.html: section 10       // real docs
  lib/.../email/google_analytics_import.html.heex // real code
  google-analytics-import.html                    // real docs
  plausible/analytics#5305: GA4 import: refresh... // real PR
  ...lib/plausible/imported/google_analytics4.ex          // real code

This wasn't hand-picked to make a chain — it's what the compiled ledger returns for one ordinary search, once a real external docs page (vendored unmodified from plausible.io/docs), real code, and real GitHub items about the same feature are all in it together. A stronger demonstration than the single chain this RFC set out to build.

Two gaps fixed before the run. One found live. One left honestly open.