[iamtoruk]

Replaces the conservative guard from #1229 with the per-provider model the review preferred but deferred.

• The problem: SessionCache.complete is one whole-cache boolean, but a scan is scoped on two axes. fix(parser): avoid marking partially scanned caches complete (#912) #1229 kept a provider-scoped run from stamping it by running a second, unfiltered discoverAllSessions() in the stamp path. On a multi-provider machine nothing else writes the flag, so scoped users never go warm: every run re-enters cold hydration and pays a full discovery across all providers (Codex reads up to 64KB per uncached rollout). That hits the Claude-scoped refresh in usage-aggregator.ts, which exists precisely to avoid touching other providers.
• Completeness now lives on ProviderSection.complete. A scoped run marks only the provider it walked; an unscoped run marks every section and the whole-cache flag, which is the only thing that can vouch for providers with no section at all.
• A section with no flag of its own falls back to the whole-cache flag, so an existing cache reads exactly as it did in both states (true = all complete, false/absent = none). The two new envelope fields are additive and ignored by older readers, so CACHE_VERSION is unchanged and nobody's session cache is thrown away. DAILY_CACHE_VERSION is untouched.
• Date axis: a run that skipped uncached sources because their mtime predates dateRange.start stores that start as the section's completeFrom floor instead of claiming all of history. A watermark rather than a plain boolean because nearly every parse is ranged - the 365-day history parse skips everything older, and withholding completeness outright would leave any machine with older sessions permanently cold. Readers pass the requested range start; a wider request than the floor re-enters cold hydration and lowers the floor as it goes.
• Discovery no longer fails open: discoverAllSessionsWithFailures reports which providers threw, and one that returned an empty list because it failed is never marked complete. Per-file and per-provider parse isolation is unchanged.
• Discovery passes in a scoped run: 2 before (one filtered, one unfiltered over every provider), 1 after. No new parse, no extra file I/O.
• Verification: npx tsc --noEmit clean; full npm test 3938 passed, 5 skipped, 0 failed across 290 files. tests/scoped-run-completeness.test.ts rewritten (dead CODEX_HOME block dropped - the codex provider resolves its dir at import time and tests/setup/env-isolation.ts already isolates it) with cases for scoped marking, three-run convergence with a cold-hydration counter, the date floor, a throwing provider, and both legacy-cache states.

Known gap: a provider skipped mid-parse on a permission error is still marked complete, as it is today. Left alone deliberately - changing it would leave a Full-Disk-Access-denied machine cold on every launch, which is a separate call.