Closes brandon-git-segment-2l6.1 (follow-up surfaced by PR #96).

What

buildRenderPayload had two failure-visibility contracts side by side: the git/cache lanes carried typed Outcomes with boundary logging, while session/today/context/metrics/tmux still .catch(() => null) and relied on interior debug() logging — two types for one behavior at one boundary ([LAW:one-type-per-behavior]).

All seven lanes now carry Outcome(ok | absent | failed); the boundary is the single log site ([LAW:single-enforcer], [LAW:no-silent-failure]); the doc comment collapsed to one sentence as the ticket specified.

Per-lane classification (where the semantics are known)

• session — getSessionUsageFromPath returns Outcome; failed parses are NOT cached in the store (only ok records enter the map — same rule as the git cache), so the next render retries.
• today — TodayInfo tightened to non-null totals ([LAW:types-are-the-program]); "no usage today" is absent; a failed seed or failed active-session ingest fails the projection loudly instead of summing a confidently-wrong total. The once-a-day seed logs its own per-session parse failures (it is its own effect edge — no render boundary to carry outcomes to).
• context — unreadable transcript is failed; transcript with no usable usage entry is absent.
• metrics — hook without a cost block is absent; MetricsInfo tightened so only lastResponseTime stays nullable (the one field whose domain genuinely has no answer sometimes).
• tmux — module cache stores the resolved Outcome; a spawn failure stays durable (no re-spawn per render) but flows to the boundary as failed instead of being demoted to null.

The deeper seam

parseJsonlFile swallowed all read errors to [] — with that in place the new failed arms were unreachable for real read errors. Now ENOENT stays [] (a new session pre-first-write is the domain's genuine "no entries"), and every other error (EACCES, EIO) propagates so the consuming provider classifies it. All three callers are the providers migrated here.

Boundary cleanup

One generic lane() helper replaces the per-lane null plumbing — the five nullP<Awaited<ReturnType<...>>>() incantations and both as Promise<Outcome<...>> casts disappeared because the types converged. One pure take() fold projects single-value lanes; the log effect happens once at the edge ([LAW:effects-at-boundaries]).

Verification

• pnpm typecheck, pnpm lint, pnpm check:protocol, pnpm build all clean
• Full suite 1133/1133; new boundary tests pin the unified contract (failed lanes log exactly once + project missing, absent lanes silent, thrown stubs totalized to failed with lane-name prefix)
• Provider tests updated to the new contract: unreadable transcript → failed (chmod 000), missing transcript → graceful zero-entry, no-cost hook → absent

Note

The z.ai reviewer account is 429-blocked (account-level, restoration requested) — the Code Review check will fail until restored; this is expected and not caused by this PR.