Observatory tracking

jcodemunch-mcp, jdocmunch-mcp, and jdatamunch-mcp are tracked weekly at https://jgravelle.github.io/jcodemunch-observatory/ alongside Express, FastAPI, Gin, Pydantic, Django, Flask, NestJS, and Cobra (11 repos total). Cron: Mondays 06:00 UTC. Config edits auto-trigger a rebuild via the workflow's push-on-paths filter.

⚠⚠ MONDAY 2026-08-31 IS THE FIRST HONEST BUILD, AND IT WILL LOOK LIKE A REGRESSION. Until v1.108.303 the pipeline cloned with --depth=1, so churn_surface read churn 1 for every file in every repo and the axis ranked nothing but complexity — every published score was flattered, ours included. Expect a step change across the roster with no upstream cause: jcm goes 81.3 (B) -> 75.6 (C), and gin's grade holds at 91.8 (A) while its churn_surface raw moves 55.45 -> 39.42 because a different symbol becomes top hotspot. Do NOT chase this as a new defect and do NOT attribute it to any release. jjg decided on 2026-08-27 to let the cron publish it rather than trigger a rebuild early. Delete this note once that build is up.

Scores are NOT transcribed here. Pull them from the live index.json before quoting them anywhere ([[feedback_verify_observatory_output]]). A transcription is always stale: the cron is Mondays 06:00 UTC, so it lags any same-week release.

As of the 2026-07-20 build, suite ranking holds: jdatamunch > jdocmunch > jcodemunch — smaller, more focused tools score higher; jcm carries more surface area. All three sit at or above the median, and all three keep their grade. Every move is <=0.4 and no grade changed, so this is drift, not signal: the generator itself advanced 1.108.114 -> 1.108.146 between crons, and the third-party repos moved on their own upstream commits. Do not attribute these deltas to any specific release without checking the scored SHA — the two 0.1s and jdoc's -0.2 sit well inside the range the generator bump alone can produce.

NestJS's persistent D is a correctness story, not a regression. The barrel-aware import graph (jcm PR #284: export * from <spec> capture + dotted-basename resolution like ./foo.service) exposed ~18 real dependency cycles in NestJS's inter-package barrel chains that its previously-broken graph hid (find_importers against injectable.decorator.ts went 0 → 791). The corrected graph keeps NestJS at the bottom because the cycles are real.

Workflow hardening: .github/workflows/observatory.yml now accepts a version workflow_dispatch input (gh workflow run observatory.yml -f version=1.93.0) to dodge PyPI CDN propagation lag, and the workdir-cache key prefix tracks INDEX_VERSION so schema bumps invalidate prior cached SQLite indexes automatically.