docs(781): re-measure §5.3 from a derived population — the zeros split four ways #800

Merged
timothy merged 1 commits from fix/781-plugin-mcp-hygiene into main 2026-08-16 10:02:16 +02:00
Owner

refs #781 (partial — see below)

Docs-only. Appends a re-measured block to docs/defect-shapes-773.md §5.3, keeping the dated 2026-08-13 table rather than overwriting it.

What the re-measurement found

Original claim Re-measured 2026-08-14
population = the listed tools Seven enabled plugins were omitted, incl. serena — the table was hand-written
811 transcripts 964 via rglob (a top-level glob sees 209 and manufactures false zeros)
mempalace dead 19 days 31 calls, last seen 2026-08-14 — active; the gap was a snapshot artifact
codex 0 skill invocations 113 codex exec calls / 17 sessions — third-heaviest tool in the corpus
LSP 0 calls Not zero — the calls are from sessions testing whether the LSP works, incl. §5.1's own probes
gitea "identical values" Not identical (user scope has "type": "stdio")

Positive control on the same query: 24,762 Bash, 5,275 Read, 4,037 Edit.

Why this is a PARTIAL, deliberately

#781's Done-when asks that every configured plugin/MCP server be "demonstrably used or removed". The evidence argues against the removals. These zeros split four ways:

  • broken — csharp-lsp MCP + two LSP servers (fixed in #777)
  • unreachable — serena: starts standalone at v1.28.1, tools never reach a session → #799
  • just enabled — three plugins turned on the same day; their zeros mean "not yet tried"
  • wrong surface measured — codex, which is in fact the corpus's third-heaviest tool

Ticking that box would mean deleting things whose zero I have just shown is not evidence of uselessness. fixes-trailer-closes-issues-that-are-not-done is the record for exactly that. So this PR lands the audit and the one supported change, and the issue stays open pending #799 and a later re-measure.

The gitea de-duplication, and why it went the other way

Resolved in gitignored local config, so it shows here only as a recorded outcome. The naive move — keep the more specific project entry — would have broken two other repos: server-management and homelab-docs have no .mcp.json and depend on the user-scope copy. The project copy was removed instead, and the survivor was smoke-tested (starts, 54 tools) rather than assumed.

refs #781 (partial — see below) Docs-only. Appends a re-measured block to `docs/defect-shapes-773.md` §5.3, keeping the dated 2026-08-13 table rather than overwriting it. ## What the re-measurement found | Original claim | Re-measured 2026-08-14 | |---|---| | population = the listed tools | **Seven enabled plugins were omitted**, incl. `serena` — the table was hand-written | | 811 transcripts | **964** via `rglob` (a top-level glob sees 209 and manufactures false zeros) | | `mempalace` dead 19 days | **31 calls, last seen 2026-08-14** — active; the gap was a snapshot artifact | | `codex` 0 skill invocations | **113 `codex exec` calls / 17 sessions** — third-heaviest tool in the corpus | | `LSP` 0 calls | Not zero — the calls are from sessions testing whether the LSP works, incl. §5.1's own probes | | `gitea` "identical values" | Not identical (user scope has `"type": "stdio"`) | Positive control on the same query: 24,762 `Bash`, 5,275 `Read`, 4,037 `Edit`. ## Why this is a PARTIAL, deliberately #781's Done-when asks that every configured plugin/MCP server be "demonstrably used or removed". **The evidence argues against the removals.** These zeros split four ways: - **broken** — `csharp-lsp` MCP + two LSP servers (fixed in #777) - **unreachable** — `serena`: starts standalone at v1.28.1, tools never reach a session → **#799** - **just enabled** — three plugins turned on the same day; their zeros mean "not yet tried" - **wrong surface measured** — `codex`, which is in fact the corpus's third-heaviest tool Ticking that box would mean deleting things whose zero I have just shown is not evidence of uselessness. `fixes-trailer-closes-issues-that-are-not-done` is the record for exactly that. So this PR lands the audit and the one supported change, and the issue stays open pending #799 and a later re-measure. ## The `gitea` de-duplication, and why it went the other way Resolved in gitignored local config, so it shows here only as a recorded outcome. The naive move — keep the more specific project entry — **would have broken two other repos**: `server-management` and `homelab-docs` have no `.mcp.json` and depend on the user-scope copy. The project copy was removed instead, and the survivor was smoke-tested (starts, 54 tools) rather than assumed.
timothy added 1 commit 2026-08-16 09:54:14 +02:00
docs(781): re-measure §5.3 from a DERIVED population — the zeros split four ways
PR Gates / CI image pin matches docker/ci (pull_request) Successful in 38s
Build ErsatzTV Image / Delimiter ban (release path) (pull_request) Successful in 44s
Build ErsatzTV Image / EF migration integrity (SQLite + MySql) (pull_request) Successful in 49s
PR Gates / Docs update reminder (pull_request) Successful in 47s
PR Gates / decisions lifecycle (pull_request) Successful in 1m1s
Build ErsatzTV Image / Functional E2E (curl + UI contracts) (pull_request) Successful in 13s
review-verdict/h10 Exempt: docs-only change (no code, no protected path)
Review verdict / Set review-verdict status (pull_request_target) Successful in 18s
Build ErsatzTV Image / API docs in sync (OpenAPI + endpoint index) (pull_request) Successful in 8s
Build ErsatzTV Image / Formatting (changed .cs conform to .editorconfig) (pull_request) Successful in 5s
Build ErsatzTV Image / Build & test (.NET) (pull_request) Successful in 9s
Build ErsatzTV Image / Build & push image (amd64) (pull_request) Skipped
PR Gates / Script tests (pytest) (pull_request) Successful in 2m55s
0874927cfc
The §5.3 table was assembled by hand. Deriving the population from the config files
instead surfaces seven enabled plugins it omits, including `serena`. An audit whose
completeness rests on remembering every entry is the failure
`testing.guard-derives-population-from-source` exists to prevent, applied to an audit
rather than to a guard.

The corpus was also undercounted: 964 transcripts via rglob, not 811 — a top-level
glob sees 209 here and manufactures false zeros. Positive control: 24,762 Bash.

Two corrections to existing rows. `mempalace` is not dead — 31 calls, last seen
2026-08-14; the 19-day gap was a snapshot artifact. `codex` is the third-heaviest
tool in the corpus at 113 `codex exec` calls across 17 sessions; the table's caveat
was right and this is the number.

The conclusion the issue expected does not survive. "Retire what is enabled and never
invoked" reads a zero as uselessness, but these zeros split four ways: broken
(csharp-lsp MCP and two LSP servers, #777), unreachable (serena, #799), just enabled
(three plugins turned on that same day), and measured on the wrong surface (codex).
Establishing WHY a counter is zero is a precondition for acting on it, and on this
evidence almost nothing should be removed. What was wrong was duplication and
breakage, not excess.

The one removal the evidence supports is `gitea`, and it went the opposite way from
the obvious choice: the PROJECT copy was dropped, not the "more specific" one, because
server-management and homelab-docs have no .mcp.json of their own and depend on the
user-scope entry. The survivor was smoke-tested afterwards (starts, 54 tools) rather
than assumed. That change is to gitignored local config, so it appears here only as
the recorded outcome.

The 2026-08-13 table is kept rather than overwritten: it is a dated measurement, and
the re-measurement is stated against it so a reader can see what moved.

refs #781

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
timothy force-pushed fix/781-plugin-mcp-hygiene from 0d792329cf to 0874927cfc 2026-08-16 09:54:14 +02:00 Compare
timothy merged commit 107716fa4f into main 2026-08-16 10:02:16 +02:00
timothy deleted branch fix/781-plugin-mcp-hygiene 2026-08-16 10:02:19 +02:00
Sign in to join this conversation.