feat(julia-ci): record the timings this workflow already reads, and stop the fallback reading as transient

[sotashimozono]

julia-ci.yml reads the ci-timings orphan branch but contains no job that writes it, so the
timings-informed LPT planning it advertises can never activate for a caller.

The header says:

timings-informed LPT planning kicks in automatically when a ci-timings orphan branch with
timings.tsv exists (else deterministic round-robin).

and the plan job implements the read (lines 60–67). Nothing in the file produces timings.tsv.
The test job does not upload per-shard timings and there is no record-timings job.

Measured, 2026-07-30

caller	shards	ci-timings branch
QAtlasHub/AbstractQAtlas.jl	4	absent
QAtlasHub/DataVault.jl	1	absent
QAtlasHub/ParamIO.jl	1	absent
lab-sotashimozono/ParaLinearAlgebra.jl (inline CI, has record-timings)	8	present, updated 2026-07-30T03:33:30Z

shards: 1 needs no timings, so DataVault and ParamIO are unaffected. AbstractQAtlas is not: it
fans out four ways and its planner has been round-robin on every run since adoption, with no path to
ever becoming anything else. The failure is silent — round-robin is a legitimate documented fallback,
so the log line "no ci-timings branch yet — round-robin fallback" looks like a first-run state
forever rather than a permanent one.

Fix

Port the producer half from ParaLinearAlgebra.jl's inline CI.yml (jobs test upload +
record-timings, lines 110–160):

• test uploads test/.ci-out/timings-*.tsv as an artifact per shard;
• a record-timings job downloads them, merges into the existing timings.tsv (latest value wins),
  and force-pushes the ci-timings orphan branch. Needs contents: write.

Gate it on inputs.shards > 1 so unsharded callers pay nothing, and on github.ref == 'refs/heads/main'
so PR runs do not race each other for the branch.

Why this blocks a larger migration

lab-sotashimozono/.github has no julia-ci, and 45 repos across the three owners still carry an
inline CI.yml (6016 lines total) that this reusable is meant to absorb — see
lab-sotashimozono/.github#15. ParaLinearAlgebra's inline version does record timings, so adopting
the reusable as it stands would be a regression for it and for every other sharded repo. Fixing the
producer here first, then porting, keeps the known defect from being copied into a second hub.

Not a gap

The compat job (minimum LTS + latest) has no counterpart here, but it does not need one:
AbstractQAtlas.jl already expresses it as a second call with julia-version: "1.10", which is the
better factoring. Only the timings producer is missing.

---

What this PR does

test step	sets CI_EMIT_TIMINGS / CI_TIMINGS_DIR (push:main and shards > 1 only) and uploads timings-*.tsv per shard
record-timings (new)	merges the artifacts into the ci-timings orphan branch, latest value wins
plan	when shards > 1 and no timing data loaded, warns that round-robin is permanent, not first-run
contract header	names CI_EMIT_TIMINGS / CI_TIMINGS_DIR and the contents: write the caller must grant

Env names are repo-agnostic on purpose. ParaLinearAlgebra's inline emit is gated on PLA_EMIT and
writes to PLA_CIOUT_DIR — names a reusable cannot set — so the neutral pair becomes part of the
documented contract alongside CI_CASES.

Failure behaviour, chosen deliberately

• record-timings is continue-on-error. ParaLinearAlgebra's inline version has already produced a
  failure verdict on main where all eight shards and the aggregate passed and only the timings
  push failed. A bookkeeping job that reddens a green suite teaches people to ignore red.
• 0 rows warns and leaves the branch alone. A suite that has not implemented the emit is a valid
  state. Pushing an empty timings.tsv would be strictly worse: plan would load zero rows and
  report success.
• The signal that replaces the silence lives in plan, which is required and runs every time. It now
  distinguishes "no data yet" from "no data, ever" — the original line read the same on run 1 and run
  500, which is why this went unnoticed across every caller.

Purely additive (84 insertions, 0 deletions). shards: 1 callers see no change.

Adoption note

lab-sotashimozono/.github has no julia-ci, and 45 repos across the three owners still carry an
inline CI.yml (6016 lines) this is meant to absorb — lab-sotashimozono/.github#15. Fixing the
producer here first keeps the defect from being copied into a second hub when that port happens.

Closes #19