Delete openrouter_pricing and the snapshot machinery it alone required

[IngTian]

Net −12,293 lines. The feed rendered on no panel, and the subsystem built to keep it
alive had just destroyed six days of it.

Why the data wasn't worth keeping

/api/v1/models reports one host's price per model, by an unstated rule, out of a
very wide distribution. Measured live on 2026-10-04 across 8 multi-hosted models:

model                        hosts      min   median      max  reported
deepseek/deepseek-v4.1-flash    31   0.0030   0.2000   0.4500   0.3000   (150x)
z-ai/glm-5.3                    40   0.0750   1.3300   2.8000   1.4000   ( 37x)
qwen/qwen3.8-27b                17   0.0240   0.2025   0.9900   0.4200   ( 41x)
deepseek/deepseek-v4-pro-0813   20   0.1760   1.2110   1.6500   0.2200   (9.4x)

median spread 20.7x · reported == cheapest host in only 3 of 8
(in 2 of 8 it matched neither min, median, mode nor max)

So a "price change" in the series was usually a change in which host got selected that
morning. It also measured one factor of a product — inference revenue is tokens × price,
and no credible free token-volume series exists, so the price half answers nothing alone.
Both recorded in sources/__init__.py so this isn't rebuilt.

And the premise was false. backfillable=False asserted the history could never be
recovered, which justified the committed CSVs, the cron and the gap test.
kj-9/openrouter-models-json has committed the same endpoint daily since 2024-10 (654
commits, pushed 2026-10-03), and the Wayback Machine has ~191 days of it back to
2023-07-26. It was archived in two other places the whole time.

Why the machinery went too

backfillable had no other use — the snapshot write, --snapshot-only, a --list column,
and one always-true guard in the row-count floor. With the only False source gone it
could only ever be True, so it is deleted rather than kept as a config value that never
varies.

The redistribution rule survives and is now stronger. It used to be "a restricted
source has no directory under data/snapshots/", which stopped meaning anything once
snapshots were gone. It is now "nothing under data/ is tracked at all" (asked of
ls-files), plus the existing assertions that no path a --db run can write to is
committable. Wider net than what it replaces: it catches a CSV export or a debug dump, not
just one known directory. Verified non-vacuous — 0 tracked files under data/ now, 27 at
HEAD.

What is left automated

The feeds job — one live verify_spec call to catch CFTC field-map drift — moved to its
own file, committing nothing, contents: read instead of contents: write. Its header
records the failure mode that killed the old workflow, because it is easy to rebuild by
accident: the commit step sat after the test step, so a dropped cron left a one-day gap,
the gap test went red, and the commit was skipped from then on.

The 27 committed CSVs are deleted from the tree and remain in history.

Verification

• 356 tests pass (from 371: 21 snapshot-specific tests removed, 10 added back)
• scripts.ingest --list → 9 sources, no openrouter; --snapshot-only now rejected
• scripts.build against a fixture copy still yields 14,079 positioning + 584
  inflation rows and reports journal_mode now delete
• scripts.status drops the snapshots section
• Both workflow files parse; no orphaned imports

Known follow-up

tests/fixtures/fixture.sqlite still carries a vestigial openrouter_pricing table. It is
simply not loaded (noted in conftest.py); the next make fixtures drops it. I did not
regenerate it here because that needs a fresh store and would churn the committed CFTC
slice the identity tests depend on.

🤖 Generated with Claude Code