test(sandbox): preserve the disposable ship target as a fixture

[schmug]

--dry-run rehearses everything except the ship. The four steps it deliberately stops before — the R2 PUTs, the feed-manifest upsert, the Pages deploy hook, the covered.json write — had only ever run against the two live shows. So the one part of the pipeline with irreversible outward effects was the one part with no safe way to exercise it.

This preserves the disposable publishing target that fixes that.

Verified end to end

Ran for real on 2026-09-10 against the real bucket:

r2_status	published, public URL HTTP 200
appeared	sandbox/<slug>.mp3 (1,896,532 B), sandbox/<slug>.jpg, manifest-sandbox.json
manifest.json	sha 74639eed… → 74639eed…, still 91 entries
manifest-frontier-commits.json	sha 26b489c6… → 26b489c6…, still 2 entries
covered.json	1064 → 1064 entries

Checked against a prediction written before the run, not rationalised after. The deploy hook fired (HTTP 200) and the probe surfaces nowhere on /podcast/, /frontier-commits/, or the real manifest.

Why the test file matters more than the fixture

The fixture is the dangerous kind of file: it points at the real bucket, it gets edited by whoever is spiking, and its entire safety property is four keys holding it away from two live feeds.

tests/test_sandbox_fixture.py re-derives the live namespaces from their own sources — the daily show's constants, and Frontier Commits' documented manifest parsed out of its SKILL.md rather than copied — and fails if the sandbox collides with either. A show that changes its namespace drags these checks with it, instead of silently leaving them protecting nothing.

Every guard is mutation-tested. Each of these fails at least one test:

mutation	caught
drop r2_key_prefix	2 failed
slug_prefix → the daily show's	1 failed
feed object → manifest.json	1 failed
key prefix → Frontier Commits'	1 failed
a segment gains a real source_url	1 failed
pin a date	1 failed
ship_mode → spotify	1 failed
drop the intro role	2 failed
music hash no longer matches	1 failed

Two deliberate choices

• No date key, so every run mints today's slug. R2 objects are immutable-cached: republishing a slug replaces the origin bytes while the edge keeps serving the old ones per POP, so a pinned date would make every later spike fight a stale cache.
• Music enabled, so one command exercises the mix and the music pre-flight too. The asset hash is pinned in the test, not only the fixture, so a re-recorded asset fails a unit test rather than a live run's pre-flight mid-spike.

Docs

CLAUDE.md gains a section on when to reach for it — a change touching the ship (feed schema, R2 keys, slug minting, the publish path) — and when not to: audio or script changes, which a dry run covers completely and where a live publish adds only litter. It also records the two things the sandbox does not isolate (the deploy hook fires; the run appends a normal web-ready record to runs.jsonl), and how to reshape it for a future show's spike.

The documented invocation was run verbatim before committing: pre-flight 8/8, artifact gate PASS, previewed URL under the sandbox prefix.

Tests

1549 passing, 0 failing. ruff check + format clean; all 7 pre-commit hooks pass.

🤖 Generated with Claude Code