fix(pr-cost): price the claude lane by model family and cache TTL, and allow a corrected figure

[cheshirecode]

What

Two pricing defects in pr-cost's claude lane, plus the missing escape hatch for republishing a figure that was priced wrong.

Why: both defects produced a plausible number, not an error

1. The rate table charged Opus 4.6+ at Opus 4.0's rates.

MODEL_RATES matches by longest prefix but carried only a claude-opus-4 row, so it swallowed claude-opus-4-6, -4-7, and -4-8. Those cost $5/$25; the row charged them $15/$75 — a 3× overestimate.

It also had no row for the current families at all, so claude-opus-5 fell through to the flat defaults. Those defaults are 5/25, which happen to equal Opus 5's real rates — so the dollar figure looked correct while usd_basis said it was not priced from the model. The dollar figure alone cannot catch that, which is why the test asserts the label.

Added: fable-5-1 (with its 0.025× cache read), fable-5, opus-5, opus-4-8, opus-4-7, opus-4-6, sonnet-5. claude-opus-4 keeps 15/75 for genuine 4.0/4.1 — longest-prefix matching separates them. sonnet-4 and haiku-4 already resolved correctly for 4-6 and 4-5.

2. Every cache write was priced at the 5-minute TTL rate.

Writes cost 1.25× input at the 5-minute TTL and 2× at the 1-hour TTL. Claude Code sessions run the 1-hour TTL. usage.cache_creation says which is which; the flat cache_creation_input_tokens total does not. This was a systematic undercharge.

The reader now reads the split and emits cache_write_5m_input_tokens / cache_write_1h_input_tokens. Transcripts predating cache_creation carry no breakdown; those writes bill at the cheaper 5-minute rate, so a missing split understates rather than inflates — and the two fields still sum to the total.

Prices verified against the bundled claude-api skill's model table and caching doc rather than written from memory.

3. A corrected figure could not be published.

The duplicate guard keys on pr_url + session_id, so the annotate carrying a corrected number for the same session was refused — the case someone hits first after learning their posted figure was priced wrong. annotate --allow-duplicate now publishes it and reports "status": "corrected", appending a second ledger row rather than editing the first, so the ledger keeps both what was published and what replaced it.

from-hook deliberately has no such flag: a hook that re-fires must stay idempotent.

Design note

Rates come from a Rates NamedTuple built by one helper that states Anthropic's published multipliers once — reads 0.1× input, writes 1.25× / 2×. A hand-typed write rate can no longer drift away from its input rate. Every pre-existing table value is reproduced exactly.

Provenance

This revives fix/pr-cost-cache-visibility (2 commits, never pushed). That branch no longer rebases onto main — main restructured skills/pr-cost/SKILL.md into a thin root plus references — so the surviving work is re-applied onto current main.

Not carried over, because main already has it: the cache read/write split in the reader (main ships it under different key names), and the dedup-by-message-id test (main has test_usage_key_contract.py).

Tests

Proved red against main:

Test	Red on main
test_opus_4_6_is_not_priced_as_opus_4_0	15.0 != 5.0
test_opus_5_is_priced_from_the_table_not_the_defaults	'cli-default' != 'model-table'
test_one_hour_cache_writes_cost_twice_input	KeyError
test_a_missing_cache_creation_breakdown_bills_the_cheaper_rate	KeyError
test_allow_duplicate_publishes_a_corrected_figure	argparse exit 2

Two are green both ways by design, and each was mutation-checked so it is not decoration:

• test_opus_4_0_keeps_its_own_higher_rate — the guard stopping the 4-6/4-7/4-8 rows from being "simplified" back into the bare prefix.
• test_from_hook_has_no_allow_duplicate_escape — fails 0 != 2 against a mutant that leaks the flag onto from-hook's subparser.

Verification

python3 -m unittest discover skills/pr-cost/tests  -> 76 tests OK
pr_cost_doctor.py --self-check                     -> OK
tests/run.sh                                       -> 157 pass, 0 fail, exit 0
ruff check (both changed files)                     -> clean

Follow-up, not touched here

pr_cost_collect.py has imported tempfile without using it since before this change (ruff F401). pr-cost is not ruff-gated in CI, so nothing fails on it today.