feat: Codex sessions as first-class sessions, with their own turns, tokens and honest cost

[daimonbot]

Closes #15

What was asked

A Codex provider, so Codex sessions are first-class sessions in atlas — their own turns, tools, tokens and cost — the way Claude sessions already are.

The transcript-side detection shipped in #14 already places a Codex node in the tree. What it places is a stub: apiCalls: 0, all-zero tokens, cost.confidence: "n/a", with the only real number stranded in reported.tokensUsed. Nothing had ever read Codex's own record of what the session did.

What changed

• src/providers/codex.mjs — new provider. discover() / buildTree() / refFromPath() over Codex's own session-metadata store, its turns/tool-calls store, and its rollout log. Read-only, no network.
• src/providers/index.mjs — a registry dispatching by ref.provider, replacing the direct claude.* calls at the 7 sites in cli.mjs and server.mjs.
• src/providers/node.mjs — the node shape, authored in one place instead of two.
• src/prices.mjs — a CODEX price table and priceCodex(), sibling to the existing Claude pricer.
• src/render.mjs, src/cli.mjs, src/server.mjs — the money-rendering fix, below.
• thread_spawn_edges is honoured for native Codex→Codex nesting. The table is empty everywhere it could be checked, so the empty case is what is actually proven; the populated case is exercised against a fixture.

A Codex session now lists, builds and renders on the same footing as a Claude one, with no provider-specific flag or route.

Cost: a correction to the issue's own premise

The issue asked for a subscription state that withholds a dollar figure. That is not what shipped, by explicit decision, because it contradicts what atlas already does: priceClaude() has never had any notion of billing mode — it is a pure function of model, timestamp and usage — and subscription-billed Claude sessions have always rendered a plain list-price figure.

The useful number is what the tokens would cost, because that is what transfers when the same workload moves to a metered account. What a flat plan happens to charge is not a property of the session.

So billing mode is not read at all, and there is no third state. Cost has the same two outcomes priceClaude() has always had: a figure flagged computed when the model has a price row, n/a when it does not.

The $0.00 fix

A cost of exactly $0.00 read as "free" where it meant "unknown". Fixed at all three sites that rendered a money value unconditionally: treeTerminal(), the list command, and indexHTML()'s session-list column. The third was not named in the issue and was found while specifying this.

Verification

• All 13 acceptance criteria pass, plus 2 contract-level checks. Verified against this host's real Codex data, not only fixtures: a live session builds with 48 API calls, 8 turns and a full token breakdown, against the stub's zeros.
• An unpriced model renders n/a on live data — one appeared mid-review and is unpriced by design.
• A host without Codex is unaffected byte for byte on every machine-readable surface: list --json, tree --json, /api/sessions and /api/tree are byte-identical to the pre-change tree, with the two scans run concurrently to avoid a false diff on a live corpus.
• node:sqlite stays behind an existence gate, so its experimental warning never reaches an existing Claude-only user — stderr is empty on such a host.
• One declared deviation: human-rendered rows do change on a no-Codex host, because that is the $0.00 fix landing — 56 of 1469 terminal rows and 43 HTML rows. Sort keys, row counts and aggregate totals are unchanged.

Follow-ups, deliberately not done here

Centralising price data into its own module, a price-refresh mechanism, per-effort price rows, and subscription-utilisation reporting. All deferred by decision; worth their own issues.

[daimonbot]

Superseded by #22, which is now merged and provides the same first-class Codex provider, provider registry, turns, token and cost ingestion, plus the React index-worker integration and the empty-corpus test isolation fix.