feat(release): publish an index of every artifact the fleet has published

[elronbandel]

Closes #519. The catalog is untouched — this publishes a second artifact beside
it, ghcr.io/exgentic/evals/index:latest.

Why

evals/catalog:latest answers "what can I launch" as a model: a family list
times an agent lineup. The lineup is copied verbatim from the workflow and never
swept, and each family is proved by probing a single agent — on the measured
assumption that gaps are whole-family, never per-agent.

#516 ended that
assumption. Measured against the live registry:

agent	published combos
claude-code, codex, gemini-cli, openclaw, opencode, pi, zerostack	759 each
openhands	1 (handbook--openhands)

So the catalog offers openhands against 99 benchmarks, 98 of which are an
ImagePullBackOff waiting for whoever picks them — the failure #81 fixed on the
family axis, reopened on the agent axis. A model of the fleet cannot be trusted;
only an enumeration can.

What it publishes

{"source": {...}, "generated": "2026-09-09T...Z", "registry": "ghcr.io/exgentic",
 "images": {"evals/handbook--openhands": ["latest"], "benchmarks/aime": ["latest"], ...},
 "labels": {"benchmarks/aime": {"eval.benchmark.description": "American Invitational…",
                                "eval.benchmark.env": "shared-env", ...}}}

Every artifact, with its tags, plus the eval.* labels each component declares.
Search and filter it; don't re-derive it.

The labels come from each component's Dockerfile: at the commit a release
publishes from, that is the source of truth the rest of the repo already treats
it as (benchmarks/RULES.md 24f makes the LABEL line canonical for
per-task-ness, and the chart's per-task.json is derived from it). Reading them
back out of each published image instead costs a manifest→index→config walk per
component for a copy that can only lag — measured at 95s against 89s, for the
same line count.

Named, then proved — because the registry can't be listed

/v2/_catalog returns all of GHCR with no way to seek to one org (last= is
rejected), and the packages REST listing is hard-capped at 10000 results — this
org is at 9900 already (1876 of them buildcache/*), so that listing is
truncating today. So names come from the tree — every component, every
benchmark × agent combo, every baked task, standalone bundles included — and each
is proved by its own tags/list. An untagged package reads as absent, since
listing one as available is how a launch becomes an ImagePullBackOff.

Measured

89s against the live registry: 8797 candidates, 6805 published (evals
5986, benchmarks 759, agents 25, core 17, models 9, gateways 3), 130 labelled
components, ~770 KB. One scoped token per 100 repositories, 32-way parallel.

Verification

• Ran end to end against the live registry four times (89–95s), valid JSON every
  time; the numbers above are from the last run.
• Both label sources measured against each other: Dockerfiles 89s, published
  image configs 95s, identical output for every component — the tree version is
  the one here.
• The openhands case above is a query over the produced index, not a claim.
• pre-commit run — ruff, ruff-format, actionlint, yaml, codespell all pass.
• No unit tests: this repo has no Python test target (tests/static is Rust +
  shell sweeps, and its rules require offline tests), and every branch here is a
  registry call. Say the word if you want a pytest target added and I'll cover
  candidates() and the failure taxonomy offline.

Rules checked against

• .agents/delivery/RULES.md 14 (fail dirty — a registry that never answered
  aborts the sweep instead of shrinking the index), 16 (it describes the
  latest channel), 17 (verified publish — the artifact is pulled back and
  cmpd, not trusted from a zero exit).
• .agents/RULES.md 8 (minimal — one new script and one step; the catalog
  path is unchanged. 277 lines, of which 187 are code: the bulk is GHCR's
  per-repository auth, the batching and parallelism that make 8797 refs take 89
  seconds, and retries that tell "never heard of it" apart from a blip).
• .agents/contributing/RULES.md 1 (resolves feat(release): publish an index of every artifact the fleet has published #519), 2 (code only), 3
  (this section), 4 (no tool attribution).

Follow-ups

• The dashboard offers pairs from the index instead of multiplying families by
  the lineup — what actually closes the openhands case.
• eval-containers list filters the index (feat(cli): list should answer for the published fleet, and --local for the checkout #514/feat(cli): list what the fleet publishes, and --local what the checkout has #515).
• When both read it, the catalog and the index become one artifact under one
  name.