CM-68342: sweep all IDEs' session context and dedup reports

[Ilanlido]

Session start now reports the device context unconditionally (previously a machine with no MCPs in the triggering IDE never created a device entity) and collects MCP configs from every registered IDE, not just the triggering one, sent as the config_files list. Unchanged payloads are skipped via a sha256 cache (~/.cycode/.session-context-cache, tenant-aware, 7-day TTL, written only after a successful send) so the common session start makes one API call instead of two.

Also: resolve plugins of git/github-sourced marketplaces through ~/.claude/plugins/cache (previously only directory-type marketplaces resolved, so their MCPs were silently missing), normalize plugin MCP content to the canonical {"mcpServers": ...} shape, and drop the legacy global_config_file field (the backend keeps its fallback for older CLIs).

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[omer-roth]

1. Cache format & compatibility: Cache stored at ~/.cycode/.session-context-cache with tenant-aware TTL. If format changes, old cache entries become harmless misses (no corruption risk).

2. Tenant ID dependency: Cache keyed by tenant_id from auth_info. If tenant ID is None/missing, caching still works but keys all nulls together. Tests verify this.

3. Plugin marketplace assumption: Assumes git/github plugins are cached at ~/.claude/plugins/cache/ (or ~/.codex/plugins/cache/). If Claude/Codex changes cache layout, plugins silently become unreachable (non-breaking, but incomplete).

4. Deterministic payload ordering: digest uses json.dumps(sort_keys=True) but config_files is constructed from dict.values() (insertion order of IDES). If IDES registration order changes (or if different Python versions/packaging cause different ordering), identical inventories could produce different digests and cause unnecessary resends or missed skips. Make the ordering explicit/deterministic before hashing (e.g., sort config_files by 'path' or include IDE name in each entry and sort by that).

5. Determinism in resolve_cached_plugin_dir: max(candidates, key=lambda d: d.stat().st_mtime) uses mtime. That is fine but could be brittle if two entries have same mtime. Acceptable, but consider stable tie-breaker (e.g., max by (mtime, name)).