feat(config): support human-readable durations with legacy compatibility

docs/adr/0002-remote-terminal-sessions.md

@@ -84,7 +84,7 @@ The agent keeps its terminal manager outside the control-WebSocket reconnect loo
 
 Closing a session must terminate the entire PTY process group, not only the shell process. Cleanup should send a hangup first, then escalate to termination and forced kill after bounded grace periods. Explicit operator close, absolute session timeout, agent process exit, and machine reboot close the session.
 
-The agent configuration sets `terminal.session_ttl_seconds` for newly created PTYs. The default is 43,200 seconds (12 hours); explicit zero disables automatic expiry. The agent owns and enforces this deadline locally because PTYs survive browser and control-plane disconnection. It advertises the policy to the control plane and includes the creation-time TTL in terminal inventory, allowing CC to mirror expiry and preserve it across CC restart. Missing fields from older agents retain the 12-hour default. Positive values are exact and values that cannot fit the platform timer are rejected rather than clamped.
+The agent configuration sets `terminal.session_ttl` (for example, `"12h"`) for newly created PTYs. The default is 12 hours; `"0s"` disables automatic expiry. Legacy `session_ttl_seconds` is accepted with a deprecation warning. The wire protocol still uses integer seconds, so fractional-second TTLs are rejected. The agent owns and enforces this deadline locally because PTYs survive browser and control-plane disconnection. It advertises the policy to the control plane and includes the creation-time TTL in terminal inventory, allowing CC to mirror expiry and preserve it across CC restart. Missing fields from older agents retain the 12-hour default. Positive values are exact and values that cannot fit the platform timer are rejected rather than clamped.
 
 An agent process restart or machine reboot still ends every terminal session because PTY file descriptors cannot be recovered by a new process. Surviving that boundary would require an external session host such as `tmux` or a separate terminal daemon and is not part of this design.
 

docs/config-durations.md

@@ -0,0 +1,74 @@
+# Configuration Durations
+
+Agent and control-plane TOML duration settings use strings with explicit units.
+Runtime code receives `std::time::Duration`; API and WebSocket numeric fields
+such as `timeout_ms`, `session_ttl_seconds`, and `latency_ms` are unchanged.
+
+## Agent
+
+```toml
+reconnect_base = "1s"
+reconnect_max = "30s"
+observation_sync_interval = "60s"
+observation_retention = "7d"
+
+[terminal]
+session_ttl = "12h"
+```
+
+`observation_retention` prunes the agent's local hotplug memory during inventory
+collection. It does not delete the control plane's last snapshot for an agent.
+`"0d"` disables this local pruning. `terminal.session_ttl = "0s"` disables PTY
+expiry. Reconnect and sync timers must be positive and fit the platform timer.
+Terminal TTL and hook retention must resolve to whole seconds because their
+protocol/storage timestamps use seconds. Reconnect and sync intervals can use
+subsecond values.
+Reconnect scheduling retains its existing 100ms minimum backoff.
+
+## Control Plane
+
+```toml
+command_timeout = "30s"
+enroll_token_ttl = "1d"
+history_retention = "30d"
+```
+
+All three settings must be positive. Enrollment TTL and history retention must
+resolve to whole seconds. Command timeout can use subsecond values.
+`history_retention` deletes old audit events and alert transitions, not device
+snapshots or saved identities. Device retention is a separate follow-up.
+
+## Legacy Files
+
+| Old Key | New Key | Numeric Legacy Unit |
+| --- | --- | --- |
+| Agent `reconnect_base_ms` | `reconnect_base` | milliseconds |
+| Agent `reconnect_max_ms` | `reconnect_max` | milliseconds |
+| Agent `observation_sync_interval_seconds` | `observation_sync_interval` | seconds |
+| Agent `observation_retention_days` | `observation_retention` | days |
+| Agent `terminal.session_ttl_seconds` | `terminal.session_ttl` | seconds |
+| CC `command_timeout_ms` | `command_timeout` | milliseconds |
+| CC `enroll_token_ttl_seconds` | `enroll_token_ttl` | seconds |
+| CC `observation_retention_seconds` | `history_retention` | seconds |
+
+Old keys accept non-negative integers in their original units, or explicit
+duration strings, and produce a warning on stderr when loaded. New keys require
+strings; bare numbers are rejected. Supplying both keys for the same setting is
+an error, even when their values agree. Negative, malformed, and overflowing
+values are rejected; zero active timers are rejected rather than silently
+clamped. No secrets are included in deprecation warnings.
+
+`init-config --from-config ...` reads legacy files and writes only canonical
+keys and formatted duration strings. Existing numeric CLI flags are unchanged
+in this slice, and their overrides also serialize in the new format.
+
+For example, legacy `reconnect_base_ms = 1500` is read as 1500 milliseconds
+and written back as `reconnect_base = "1s 500ms"`. New-key
+`reconnect_base = 1500` is rejected: no implicit numeric unit is assigned.
+Serialization always writes duration strings, regardless of the accepted input
+format, using `humantime-serde`.
+
+The parser is [humantime](https://docs.rs/humantime/latest/humantime/fn.parse_duration.html).
+Examples include `"14d"`, `"168h"`, `"250ms"`, and `"1h 30m"`. These are elapsed
+durations, not calendar scheduling rules; use days/hours for retention rather
+than assuming calendar-month semantics.

docs/roadmap.md

@@ -22,6 +22,8 @@ live browser recovery and shutdown behavior are not established by those checks.
 
 ## Device Management And Presentation
 
+- [x] Endpoint-derived presence on snapshot writes and fleet reads; endpoint-less legacy rows retain their summary. CLI ordering matches the core signal order.
+
 - [x] Make **forget remembered device** reachable in the frontend, with confirmation. The existing backend removes the known identity and its identifiers, not agent observations. An observed device may reappear as unknown.
 - [x] Add remembered-device rename support, with backend validation and frontend editing (`PATCH /api/v1/control/devices/{device_id}`). Pinned state and identifiers are preserved.
 - [ ] Check and complete editing of remembered-device notes.
@@ -34,6 +36,8 @@ of remembered identifiers based on observation expiry.
 
 ## Telemetry Follow-Ups
 
+- [x] Human-readable agent/CC config durations with unit-aware legacy loading, stderr deprecation warnings, and canonical init-config output. See [configuration durations](config-durations.md). History retention remains separate from future device expiry.
+
 - [x] Separate console, OTLP log, and trace filters using per-layer filtering. Explicit signal config overrides `RUST_LOG`, then CLI verbosity supplies defaults. Unconfigured defaults remain info-level; exporter-failure logs stay local. Live validation pending.
 - [ ] Record hostname-resolution success/failure without exporting hostnames or resolved addresses. Current spans show timing but do not explicitly classify the result.
 - [ ] Investigate the slow resolver with a reproducible device-side check. Timing isolated the phase, not the cause; do not treat instrumentation as a fix.