feat(intent): maintainer-requested PR-channel cap (~10 open) + two friction findings

_kos/findings/finding-021-just-check-pyyaml-env-gap.md

@@ -0,0 +1,33 @@
+# finding-021 — `just check` assumes PyYAML in system python; fails wholesale on clean machines
+
+**Date:** 2026-07-24 · **Run context:** run-17 wrap (manifest v0.5 edit) · **Class:** tooling friction (capture-before-workaround per aae-orc `.claude/rules/tooling-friction.md`)
+
+## Symptom
+
+`just check` exits 1 on mokuzai with `ModuleNotFoundError: No module named 'yaml'`
+for **all five** YAML files (pack.yaml, three intent manifests, _kos/kos.yaml) —
+including files untouched by the session. The recipe (justfile line 14–15) shells
+out to bare `python3 -c "import yaml; ..."` per file; nothing declares or
+bootstraps the PyYAML dependency. Any machine whose system python lacks PyYAML
+fails validation regardless of manifest correctness. `just lint` degrades
+gracefully on the same machine (markdownlint/yamllint probe with `command -v`
+and skip); `check` does not.
+
+## Workaround applied (this session)
+
+Ran the same per-file validation through `uvx --with pyyaml python3 -c ...`
+(uv-managed ephemeral env, per the global uv/uvx preference) — all five files
+pass, including the new v0.5 `maintainer_capacity.pr_channel` block.
+
+## Why this is happening
+
+The recipe was authored on a machine where system python had PyYAML (brew
+python site-packages, most likely). The dependency was invisible at authoring
+time — classic works-on-my-machine env coupling in a repo whose CI (if/when it
+runs `just check`) would install deps explicitly.
+
+## Fix shape (not applied here — one-line recipe change)
+
+Point the recipe at uv: `uvx --with pyyaml python3 -c ...` (or probe like the
+lint recipe does and emit an actionable skip). Cheap, bounded; belongs in a
+hygiene commit touching only the justfile.

_kos/findings/finding-022-github-ssh-endpoint-intermittent-publickey.md

@@ -0,0 +1,37 @@
+# finding-022 — GitHub SSH auth fails intermittently per backend endpoint; identical key accepted seconds later
+
+**Date:** 2026-07-24 · **Run context:** run-17 wrap (pushing feat/pr-channel-cap) · **Class:** tooling friction (capture per aae-orc `.claude/rules/tooling-friction.md`)
+
+## Symptom
+
+`git push` / `git ls-remote` / bare `ssh -T git@github.com` all returned
+`git@github.com: Permission denied (publickey)` for ~5 minutes, offering the
+correct, long-registered key (`~/.ssh/id_ed25519_github`,
+SHA256:HkD3evhf2qSlI9OauLLizAtqkOfeumHpTS70g2mBWos, authenticates as
+`arcavenai`). The SAME command with the SAME key succeeded immediately before
+and after the window. Verbose transcripts isolated the variable: success
+authenticated against `140.82.112.3`, rejections came from `140.82.114.4`.
+Server host key verified as GitHub's genuine ed25519 both times — no MITM, no
+local config change, agent state irrelevant (key is `IdentitiesOnly` from
+disk).
+
+## Diagnosis
+
+Endpoint-dependent auth flakiness on GitHub's SSH front-end pool (backend
+`18bbdcb`). Nothing on this machine was wrong.
+
+## Workaround / correct response
+
+**Retry.** Each attempt may land on a different backend; a 4×/3s retry loop
+succeeded on attempt 1. Fallbacks if a bad backend persists: HTTPS push with
+`-c credential.helper='!gh auth git-credential'`, or SSH over
+`ssh.github.com:443` (different pool).
+
+## Anti-pattern this finding exists to prevent
+
+`Permission denied (publickey)` reads as a local key/agent/config problem, and
+the obvious moves — re-adding agent identities, editing `~/.ssh/config`,
+regenerating keys, touching `core.sshCommand` — are all wrong here and leave
+config debris. Before rewiring anything, check whether the failure is
+endpoint-intermittent: run `ssh -Tv git@github.com` twice and compare the
+connected IP and result.

skills/beadle-triage/SKILL.md

@@ -319,7 +319,11 @@ cadence, and calling rational scarce-attention triage "drift."
    not engagement. Do not grow the open measured-issue count while the
    acted-share is below the manifest threshold; prefer fix-PRs-attached
    (empirically the entire selection function); zero-new-issue pauses while
-   the maintainer drains are correct behavior, not idleness.
+   the maintainer drains are correct behavior, not idleness. The same
+   governance applies per channel: honor any manifest-declared channel cap
+   (`maintainer_capacity.pr_channel` — e.g. vsdd-factory's maintainer-requested
+   ~10 concurrent open measured PRs); at the cap, queue locally and stagger
+   rather than opening more.
 7. **A4 discharge path is the measured side's.** The alarm stays (data-loss
    classes are why beadle exists), but its designed discharge is lowering
    activation energy — shipping SDL fix PRs — not asking part-time humans to

targets/vsdd-factory.intent.yaml

@@ -8,7 +8,11 @@
 #       _kos/ideas/capacity-adjusted-direction-baselines.md and SKILL §6b).
 #       Direction verdicts, streak denominators, success bands, and prediction
 #       horizons are conditioned on this block. Operator-ratified 2026-07-22.
-schema_version: 0.4
+# v0.5: added maintainer_capacity.pr_channel — maintainer-requested soft cap on
+#       concurrent open measured-side PRs (~10; Zious11 via operator, 2026-07-24:
+#       the ~25-PR backlog preceding the 2026-07-19/22 drains was too much for
+#       his analysis tooling). SKILL §6b rule 6 governance, PR-channel analog.
+schema_version: 0.5
 
 target:
   repo: drbothen/vsdd-factory
@@ -46,7 +50,7 @@ maintainer_capacity:
   context: one of hundreds of projects for these maintainers; primary focus elsewhere
   attention_window: action-day   # a calendar day with >=1 maintainer action on this target
   # Derived from the store's maintainer_actions dates; refresh each run.
-  observed_windows: ["2026-07-08", "2026-07-15", "2026-07-19", "2026-07-21", "2026-07-22"]
+  observed_windows: ["2026-07-08", "2026-07-15", "2026-07-19", "2026-07-21", "2026-07-22", "2026-07-23"]
   success_bands:
     pr_acceptance:               # vs human ~68-73% / bot ~37-45% literature bands
       green: ">= 0.60"
@@ -60,6 +64,14 @@ maintainer_capacity:
   # All streak metrics (A4 included) report attention-window denominators FIRST;
   # runs/calendar as shadow. Pre-registered prediction horizons are denominated
   # in windows (default 4-6) and must test single-window actions (SKILL §6b).
+  # Maintainer-requested PR-channel cap (Zious11 via operator, 2026-07-24):
+  # keep concurrent open measured-side PRs to ~10. The ~25-PR backlog before
+  # the 2026-07-19/22 drains overloaded his analysis tooling. Soft cap, same
+  # governance class as §6b rule 6 (filing-rate), applied to the PR channel:
+  # at the cap, queue fixes locally and stagger — do not open more.
+  pr_channel:
+    max_open_measured_prs: 10
+    source: "maintainer request, relayed 2026-07-24 (Zious11 — analysis tooling)"
 
 # Composite, weighted intent anchor (no single source is authoritative).
 intent_sources: