fix(watchdog): skip cron-liveness rescue when daily-regen is disabled - #10540

#10628)

## Why

The `babysit-pipeline` skill covers watching one fresh spec to 15/15.
Backfilling coverage gaps across the catalogue — dozens of older specs
each missing the libraries added after they were generated — is a
different job, and the skill said nothing about it. Running it by hand
on 2026-08-24 surfaced four traps that cost, or nearly cost, real
implementations.

## What's new

**Section 5 · Gap backfill**

- Compute the missing set from `plots/{spec}/metadata/{lang}/{lib}.yaml`
on `origin/main`, **not** from `impl:{lib}:done` labels — most affected
specs have closed issues, so their labels are absent or stale. Includes
the `git ls-tree | awk` one-liner.
- `run_spec.sh <spec> <model> <lib>...` as the driver: staggered
dispatch, poll to metadata, `RESULT=COMPLETE|PARTIAL|TIMEOUT`, skips
libraries already on main so a re-run retries exactly the gaps.
- Two specs in parallel (~4 specs/h vs ~2; ten concurrent
`impl-generate` runs showed no rate-limit effects).
- A `done.log` / `deferred.log` ledger written *before* dispatching the
next spec, so a compaction or crashed session resumes without
recounting.
- **The one-retry rule**, with the evidence for why it is not optional:
the single retry recovered `highcharts/treemap-basic`,
`ggplot2/wireframe-3d-basic` and `ggplot2/network-force-directed`, all
three of which had already been read as capability gaps.

**Four new gotchas**

| Trap | Why it matters |
|---|---|
| `Marking <lib> as failed: N generation attempts` counts more than this
run | The markers span past campaigns (#10627 scopes this to 12 h).
Check the `Previous failures for <lib>/<spec>: N` notice before
concluding a library can't do a plot type. |
| `impl:<lib>:failed` is terminal **and** unreliable | Nothing
re-dispatches it — watchdog case 3 fires once, then only logs `needs
manual attention`. And of 87 such labels, **42 sat on implementations
that had since landed**. |
| "Agent reports success, writes no file" | 8 of 85 generate runs (~9%).
Self-heals when retry budget remains (`bubble-basic/highcharts`: failed
17:55, succeeded on auto-retry 18:02). One occurrence is noise. |
| Static library + interactive/3D spec | The one gap shape that usually
is genuine — 18 of the 45 real gaps. Still give the one retry, then
defer. |

**`run_spec.sh` joins the bundled scripts.** Its hardcoded
`REPO=/home/tirao/anyplot` is replaced by resolution from the script's
own location (`ANYPLOT_REPO` overrides), matching how `poll_spec.sh`
derives `HERE`.

## Verification

- `bash -n` on `run_spec.sh`; repo auto-resolution checked from the
skill directory (`git -C "$HERE" rev-parse --show-toplevel` → the repo
root).
- The skill's existing `ALL_LIBS` gotcha now names `run_spec.sh`'s
`lang_of` too, so a library addition updates both scripts.
- Content is drawn from a real backfill run, not invented: every number
above has a corresponding workflow run or label query behind it.

## Related

- #10627 fixes the retry-cap bug the first gotcha describes.
- #10540 fixes the watchdog crash found in the same session.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_01RbZuWNDFy7kjXh9kLfA4dP

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

…er's fetch (#10660)

## Why

#10628 documented gap backfill while the backfill was still running.
Finishing it falsified three of its claims, so this PR corrects them
rather than leaving confident wrong guidance in a skill. Running the
backfill in two parallel slots also exposed a race in the driver, fixed
here.

**Scope: prose plus one executable change.** `SKILL.md` and
`CHANGELOG.md` are documentation; `run_spec.sh` gains a fetch retry
(behaviour change, details below).

## Corrections

**1. Halt-on-cluster counted the wrong thing.** The threshold was "≥5
failed `Generate:` runs in minutes = quota exhausted". Since #10627
restored the full retry budget, a single impossible pair spends three
runs on its own auto-retries — so two bad pairs trip a raw count of five
while the pipeline is perfectly healthy. That happened during the run: 6
failures at 22:29–22:37, all of them `plotnine` on two specs, nothing
wrong with the pipeline. Now counts distinct `(spec, library)` pairs.

**2. "Static library + interactive/3D spec is the one gap that is
usually real" was backwards.** Every category-level prediction made
during the backfill turned out wrong:

| Prediction | Outcome |
|---|---|
| chartjs can't do treemap / sankey | chartjs succeeded on
`network-force-directed`, `arc-basic`, and every JS-block spec |
| plotnine can't do 3D | `bar-3d-categorical` succeeded; `scatter-3d`
did not |
| ggplot2 can't do wireframe | succeeded on retry |
| pygal is a lost cause (6 failed labels) | succeeded on
`map-marker-clustered` and `line-stress-strain` |

17 of 20 parked pairs generated fine. Exactly three failed under a full
budget — plotnine on `scatter-3d`, `contour-3d`, `line-3d-trajectory`,
all of which need a spatial projection plotnine does not have, while the
"3D" spec representable in 2D went through. The gotcha now says:
measure, don't predict.

**3. The one-retry rule needed a precondition.** The workflow now spends
three attempts per campaign by itself, so a pair that comes back missing
may already be measured and the manual retry adds nothing. The skill now
shows the actual command — `gh run list` has no per-spec filter, so it
filters the output by run title with jq. Three failures minutes apart is
a gap, one is a flake. The same check resolves `RESULT=TIMEOUT` with
`recent generate failures: 0`, which only means the failures aged out of
the driver's 25-minute window (seen on `line-3d-trajectory`, which had
in fact failed three times 40 minutes earlier).

## New gotcha

Spec IDs harvested from `impl:*:failed` issue titles must be intersected
with the real `plots/` directories: **14 of 26** pointed at specs that
no longer exist on main, and such a dispatch dies seconds in at
`Validate specification exists`. This is how the rescue list first read
as 45 missing implementations when only 20 were real.

## Behaviour change: `run_spec.sh` retries its fetch

Two drivers polling the same checkout collide on the ref lock:

```
error: cannot lock ref 'refs/remotes/origin/main': is at 6b666c1 but expected c736a84
```

A lost fetch leaves `origin/main` stale, so `meta_present` understates
what has landed and the poller reports `PARTIAL` for libraries that are
already merged — a failure that disguises itself as a stalled spec.
`fetch_main()` now retries three times with backoff and, when all three
lose, logs the warning **with git's own stderr** so a ref-lock collision
is distinguishable from an auth or network failure.

## Verification

- `bash -n` on `run_spec.sh`; the failure path exercised against a bad
ref → `last error: fatal: couldn't find remote ref …`, confirming the
message survives to the log.
- Every number in the corrections traces to a workflow run or label
query from the 2026-08-24 backfill; the ledger is in
`agentic/runs/babysit-2026-08-20/`.

## Related

- #10627 (retry cap) — verified in production: `Previous failures for
plotnine/contour-3d since 2026-08-24T10:38:22Z: 2` followed by `3 failed
attempt(s) in the last 12h (cap: 3 per campaign)`.
- #10628 — the section this corrects.
- #10540 — verified by dry run: `daily-regen liveness: workflow state is
'disabled_manually', not active — rescue skipped`, scan completing
successfully.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_01RbZuWNDFy7kjXh9kLfA4dP

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>