fix(go): close all medium-divergence parity findings (M1-M19)

[rado0x54]

Closes every remaining medium item from the parity audit (data/go-parity-findings-2026-07-04.md). M2/M4/M11/M12/M17 were already closed on go-parity-highs; this branch lands the other 14, one commit per logical group.

Session & transport semantics

• M6 — post-mortem sessions: transport-driven closes now retain the session (registry + buffer) so tail/re-attach show the final output; lists hide closed but show error (exact Node listSessions filter, including the close-on-closed no-op). Only an explicit Close clears the buffer and removes the entry. The MCP cap prune treats retained dead sessions as dead.
• M7 — transport errors: the SSH pump distinguishes a broken transport (reset, protocol error) from a normal end (clean exit, nonzero exit, hangup without exit-status) — errors now surface as status=error / transport-error in WS + audit.
• M8/M9 — Denying one WebAuthn key aborts SSH connection instead of trying next identity #91 regression: the terminal path now cancels a dead connection's stranded sign prompts (OnConnectionEnded hook on the passkey factory, fired on connect failure and transport end) and Broker.CancelForConnection broadcasts sign:resolved so stale toasts clear on every tab (agent-proxy path switched to it too).
• M10 — approval store pruning: terminal-state actions are deleted 120s after expiry (Node's poll-grace window); the store no longer grows forever.

MCP surface

• M3 — server instructions: live endpoint list + workflow/session-lifecycle/sudo guidance, byte-matching Node's text; serverInfo.version now comes from build info.
• M5 — per-account cap: the MCP path resolves accounts.max_sessions instead of always defaulting to 5.
• M1 — notifications: closed as SDK-blocked. The official go-sdk (v1.6.1; unchanged in v1.7.0-pre.1) has no public API to send custom notification methods — notifications/shellwatch/* is rejected by its typed method registry and the underlying jsonrpc2 connection is unexported. Until the sdk grows an escape hatch, agents must poll; the instructions deliberately omit Node's notifications section so they poll instead of waiting for notifications that cannot arrive. Recorded as a known divergence.

Wire & config

• M13: first-party/skip consent auto-accept sends remember: true, remember_for: 30d — the SPA client shows up in "Authorized clients" and remembered skips renew.
• M14: endpoint delete is a soft delete (enabled=0) and lists filter enabled — Node parity including the get-by-id quirk (post-delete PUT/session-create by id still work); Node-era deletions can no longer resurrect on a shared data dir.
• M15: action view createdAt/expiresAt are epoch-ms integers per the contract.
• M16: audit pages serialize "rows": [] when empty, page cap is 500, and the sessions cursor payload uses sessionId — pre-migration cursors keep working.
• M18: an absolute keyDirectory is kept as-is (path.resolve semantics).
• M19: seeding failure aborts boot — a fresh deployment can't come up silently un-loginable.

Tests

New coverage: manager retention semantics (post-mortem read, list filters, no-op close), server-hangup test updated to pin retention, MCP instructions assertion, config/store behavior via the existing suites. go build && go vet && go test -count=1 ./... and pnpm spdx:check clean.