[gregoryfoster]

Closes #138. PR 2 of 2: the content.capture consumer, on PR 1's runtime (#157). Off by default (REPLICATOR_CAPTURE_ENABLED); merge after runbook phase A is deployed and read back.

• Consumer: CAPTURE_SPEC (group replicator.capture, dedupe segment capture), capture.py (validate → pace → render → store → capture_stored) and capture_reporter.py, both facts on content.artifacts. No content.blobs fact, no change fingerprint.
• Boot: the group is created at 0, never $; one test launch; a browser that can't start degrades capture alone (worker ready: capture: unavailable, a CRITICAL naming the runbook).
• One HostPacer for fetch and capture, lifted into main.run; pace re-reads after each wait, so two loops on one host go out an interval apart (CR 14).
• Settings: the flag, group/consumer, REPLICATOR_CAPTURE_MAX_WAIT_SECONDS (3 × wait + 30 ≤ REPLICATOR_MAX_FETCH_SECONDS, CR 5), …_MEMORY_MAX_BYTES, …_CGROUP, REPLICATOR_PLAYWRIGHT_BROWSERS_PATH.
• Contract: docs/contracts/content-capture-issuer-contract.md, the normative reason registry (invalid_request_options, wait_timeout (terminal), destination_refused, not_fetchable, http_status, too_large, plus the loop's two). The charter moves archival capture out of future; capture_reporter.py joins DOMAIN_ECHO_MODULES.
• Docs: AGENTS, STREAMS, COMMANDS (four streams, four DLQs), ENVIRONMENT, ARCHITECTURE, POLITENESS, CONVENTIONS, README; runbook phase B; CI's import smoke names the capture types.
• Tests: scripts/seed_capture.py (scratch only); the loop against a scratch broker; the real shell through the handler to capture_stored.

Its final tree is the one reviewed (CR 1–36 on #138) and tested: full suite 1671 passed, -m browser 23/23, scratch-broker integration 15/15.

Post-merge host steps: runbook phase B (deploy still off → the operator writes the flag → worker ready reads capture: enabled → broker#116's acladmin read-back → archiver#380 may issue).

🤖 Generated with Claude Code