feat(projects): per-repo cost, and the mixed-era rule that makes it safe

[pitimon]

Summary

Closes the cost half of #102. The surfacing half shipped in #118.

Project rows were keyed project|source|hour and carried no model, so per-repo cost was not merely unimplemented — it was absent from the data model. The README's "which repo, which model, and which hour" promised a repo × model join that had never existed.

The key is now project|source|model|hour and the row carries the model, so computeRowCost — the same function the rest of the product prices with — works per project. A repo total and a model total cannot disagree, because they come from one function.

The transition is the whole job

This is the part that can be wrong in a way nobody notices, and it is what #102 flags.

A legacy row (project|source|hour) and a new row (project|source|model|hour) describe the same bucket at different granularity. Keep both → double-count. Drop either → lose real usage. Neither is acceptable, and neither is what happens here.

The bucket state in cursors.json is re-keyed on load. An old bucket becomes the unknown slot for its hour, keeping its running total and its queuedKey. It then stops receiving new usage — that goes to the model-specific bucket — so it is frozen at its pre-upgrade total and the two sum to the true number. The reader keys model-less rows to that same unknown slot, so a legacy row and its re-keyed bucket are one identity, not two.

Without the re-key the old bucket strands: its key never matches again, new usage restarts from zero under the model key, and the reader still sees the old row. That is the double count, and it is why this is a state migration rather than a key change.

The migration runs in normalizeProjectState, which every parser goes through, so one call covers all seven. Idempotent — a four-segment key is skipped — and it refuses to clobber an existing unattributed bucket, leaving the legacy one in place for a human rather than silently discarding either.

Cost is honest about what it cannot price

A row with no model prices at 0 and is counted as unattributed_tokens rather than folded into the total.

Repo	Shows
all rows carry a model	$3.00
all rows predate the change	no model recorded — not $0.00
mixed	$3.00+
server sent no cost field at all	nothing — absent is not zero

Pricing model-less rows at zero would render "recorded before we recorded models" as "this cost nothing" — exactly the tier #94 was added to avoid.

Test plan

• ci:local exit 0 — 948 root tests (+15), 266 dashboard (+4)
• Mixed-era, the risky part: a legacy row and a model row for one bucket count exactly once each (700 + 300 = 1000); repeated appends of one key still collapse to the last; two models in one hour are two slots, not one overwriting the other; legacy rows repeat-append and still collapse
• Migration: a stranded bucket is re-keyed, keeps its running total and its queuedKey (without which the row re-appends), running it twice changes nothing, a model-keyed bucket is left alone, an existing unattributed bucket is not clobbered, and garbage state does not throw
• Cost: priced from the model on the row; a legacy row is unattributed rather than a confident $0 and its tokens are still counted; a mixed repo reports the priced part and names the rest
• Rendering: four component tests including "server sent no cost" → no line at all
• Every touched function under the 50-line rule

Not in this PR

Backfilling model onto historical rows — out of scope per the issue: the source logs may no longer exist, and unattributed is the honest rendering of "we recorded this before we recorded models".