fix(supervise): price a bridge turn that carried no provider receipt

[drewstone]

Problem

A cli-bridge turn whose provider reported no billed dollars reached the dollar channel as a known $0 (supervise/runtime.ts, the yield { kind: 'cost', usd: 0, usdKnown: false } at turn settle).

Measured: across 292 fleet runs under discovery-lab/pursuits, spentTotal.usd > 0 on zero runs. The dollar channel reported nothing on runs that certainly spent money.

Pairs with drewstone/cli-bridge#159, which forwards claude's total_cost_usd as a real provider receipt. This PR covers everything that has no receipt to forward.

Change

Price such a turn from the model catalog against that turn's own token counts, using the response model the bridge already sends on every chunk. Nothing new on the wire.

New src/runtime/supervise/cost-estimate.ts — priceUnreceiptedWork({ inputTokens, outputTokens, model }).

A receipt and an estimate must not read the same, so the priced part is carried explicitly:

Field	Meaning
Spend.usd	all dollars attributed
Spend.usdEstimated	the part priced from the catalog
usd - usdEstimated	what a provider is known to have billed
usdKnown	whether the total is a measurement at all

usdEstimated is added to the cost UsageEvent and to Spend, and folded through every merge site (spendFromUsageEvents, foldUsage, spawn-journal clone/add, trajectory clone/add, supervisor accumulate/add). usdEstimatedOf in runtime/util.ts returns nothing when no input carried one, so a pure-receipt fold never gains a usdEstimated: 0 that would read as "checked, none estimated".

assertValidSpend refuses an estimated part that is negative, non-finite, exceeds usd, or claims usdKnown: true.

Rules that keep this honest

• Only a turn that billed NOTHING is priced. A turn holding a partial receipt already put real dollars on the channel; a whole-turn catalog price on top would charge the same tokens twice.
• An unpriced model contributes no dollars and leaves the turn usdKnown: false — unknown, never free.
• Cache reads are priced at the full input rate. The catalog entry is { input, output } per model with no cache-read rate (verified on the published @tangle-network/agent-eval@0.145.2 .d.ts). That overstates a cache-heavy turn, which is the correct direction — an invented discount would understate spend. Asserted in the tests rather than left as a comment.
• A dollar cap is unaffected. observe and reconcile still refuse unknown dollars under a maxUsd root, and an estimate rides usdKnown: false. Test included.

Proof

tests/kernel/dollar-estimate.test.ts — 14 new tests.

Behavioral proof on the bridge path: stash src/runtime/supervise/runtime.ts, run bridge-executor.test.ts:

 × prices a turn the bridge reported no price for, and keeps the dollars unknown
 × keeps dollar cost unknown when a later completed turn omits price
 × interrupts an active response body, accounts its partial usage, and resumes with the steer
AssertionError: expected [ …(3) ] to deep equally contain { kind: 'cost', usd: 0.0000062, …(2) }
 Tests  3 failed | 28 passed (31)

With the change:

 Test Files  1 passed (1)
      Tests  31 passed (31)

Full suite:

 Test Files  214 passed | 2 skipped (216)
      Tests  2669 passed | 6 skipped (2675)
   Duration  32.65s

pnpm typecheck and pnpm lint clean.

What I did NOT change, and why

routerToolsInlineExecutor (which chatTransportExecutor builds on) has the same "no receipt → $0" shape, and I initially fixed it too. tests/kernel/chat-transport-executor.test.ts then failed on a test named "marks tokensKnown/usdKnown false when the transport omits usage or cost — never estimates", whose inline comment reads "the priced model id must NOT tempt a local estimate."

That is a deliberate, documented contract in the opposite direction, written for a bare caller-supplied transport. I reverted that half rather than flip another author's stated invariant unasked. Flagging it for a decision instead.

Docs

docs/architecture.md gains the dollar-channel section next to the token-channel rules: receipt vs price, what usdEstimated means, the cache-rate gap, and the unchanged cap behavior.

[tangletools]

✅ Auto-approved drewstone PR — bc2f0adc

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

This approval is provisional. It rests on the audit running. If the audit cannot run — for example the CLI bridge rejects it — this approval is dismissed rather than left standing, so an unrun check never reads as a passing one.

_{tangletools · auto-approval · reason: drewstone_author · 2026-08-14T05:36:29Z}

[tangletools]

✅ Auto-approved drewstone PR — bc2f0adc

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

This approval is provisional. It rests on the audit running. If the audit cannot run — for example the CLI bridge rejects it — this approval is dismissed rather than left standing, so an unrun check never reads as a passing one.

_{tangletools · auto-approval · reason: drewstone_author · 2026-08-14T05:37:41Z}

[tangletools]

✅ Auto-approved drewstone PR — 4cc0f3f8

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

This approval is provisional. It rests on the audit running. If the audit cannot run — for example the CLI bridge rejects it — this approval is dismissed rather than left standing, so an unrun check never reads as a passing one.

_{tangletools · auto-approval · reason: drewstone_author · 2026-08-14T05:55:20Z}

[tangletools]

✅ Auto-approved drewstone PR — 8a1c4322

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

This approval is provisional. It rests on the audit running. If the audit cannot run — for example the CLI bridge rejects it — this approval is dismissed rather than left standing, so an unrun check never reads as a passing one.

_{tangletools · auto-approval · reason: drewstone_author · 2026-08-14T05:57:44Z}