signerd: durable daily-cap — reserve-before-sign + forward-only rollover + crash-budget (#108) - #125

Summary

Makes the signer daemon's daily spend cap durable (issue #108, reviewed + re-scoped to "Option A" via /autoplan). The cap (spent_today_wei) was in-memory and force-zeroed on every daemon load, so any restart — crash, OOM, app update, or a same-uid attacker crash-looping the auto-respawning daemon — silently reset the day's accounting and re-opened the within-cap drain. This persists the counter and reserves the spend before signing: a new spend.json ({chain_id, account, day, committed_wei, reserved_wei}) written with the same atomic recipe as the vault, recovered into the policy on boot. A crash between reserve and commit is counted as spent on reboot (conservative; recovery only ever tightens the cap). The UTC-day rollover is now forward-only (a backward wall-clock can't reset the window), and a non-resetting supervisor crash-budget stops respawning a crash-looping daemon instead of hot-looping at the 200ms backoff floor.

Scope was deliberately held: no nonce-keyed RPC reconciliation (that restructures the frozen broadcast path to defend a sub-second window the file-delete attacker bypasses anyway), and the security value sequences with #72 (the cap values still live in an unauthenticated policy.json). See the issue and its /autoplan review comment for the full reasoning.

Key changes: extract deckard_core::atomic_write(path, &[u8]) from Vault::write_atomic (reused for the counter); new deckard-signerd/src/spend_store.rs; reserve/commit/release wired into Daemon::execute (clean RPC rejection releases the reservation, a timeout keeps it counted); forward-only rollover; account binding at unlock; supervisor crash-budget; new reserve_failed deny tag.

Linked issue

Closes #108

Definition of Done

• cargo fmt --all --check is clean
• just check is green — clippy -D warnings on the default config, --features tray, and the dev-signerd-bin arm
• cargo test --workspace is green (incl. 11 new spend_store unit tests, 2 supervisor crash-budget tests, and 2 anvil e2e durability tests)
• No new or changed dependencies (Cargo.toml / Cargo.lock untouched — only std, and already-present serde/serde_json/alloy-primitives/anyhow)
• No visual/UI change (signerd daemon internals only)
• No secret is logged or Debug-printed (the counter holds only chain id, account address, day, and wei amounts — never key material)

$ cargo fmt --all --check
# (exit 0, clean)

$ just check
# (clippy -D warnings green on workspace, -p deckard-app --features tray, -p deckard-signerd --features dev-signerd-bin)

$ cargo test --workspace
# (green — see below)

Notes for reviewers

• Security trade-off (documented): the counter file is itself same-uid-rollback-able. That residual is accepted (ADR 0004) — the durable counter's real value is honest-restart durability; the crash-loop drain is additionally killed at the mechanism by the supervisor crash-budget, and the security value completes when Spike: make policy.json tamper-evident + define a secure edit/maintenance path (not in the vault) #72 authenticates the cap values. The reframe (HIGH-security → durability) and the dropped nonce-reconciliation are explained in the issue's /autoplan comment.
• F4 (the subtle one): a broadcast timeout keeps the reservation counted (the tx may have landed — releasing would re-open the double-spend); only a clean RPC rejection releases it.
• fsync under the daemon mutex: the reserve/commit writes land on the lock that also serves STOP. Accepted for v1 (anvil-instant), matching the existing "lock held across broadcast" precedent; moving them off-lock via spawn_blocking is a noted fast-follow if the latency bites.
• Deferred (out of scope, by review): nonce-keyed exact reconciliation (post-Spike: make policy.json tamper-evident + define a secure edit/maintenance path (not in the vault) #72 issue); policy.json authentication (Spike: make policy.json tamper-evident + define a secure edit/maintenance path (not in the vault) #72); the app-side UI surfacing of DaemonSupervisor::is_crashed_out() (the flag is exposed; the banner is a follow-up so this PR doesn't touch the GPUI app).
• Test caveat: the 2 anvil e2e tests skip gracefully when anvil isn't on PATH. The conservative crash-restart guarantee (orphaned reserve counts as spent) is proven at the store level (reserved_leftover_on_reload_counts_as_spent) rather than with a process-abort fault hook.