chore: first-run telemetry and lazy-install follow-ups from the v0.11.1 release review

[anandgupta42]

Small follow-ups collected during the v0.11.1 release review that are each too minor for their own issue. Deferred because none affects correctness for this patch.

• Worker stall events carry session_id: "": startLoopMonitor() runs in cli/tui/worker.ts before setContext(); early-boot stalls lose session correlation. Backfill once setContext runs or use a boot placeholder. (CTO)
• CLI_COMMAND_NAMES in src/index.ts is hand-maintained with no test that it matches the commands registered with yargs; a new command silently falls through to "tui". Add a test diffing the set against the built yargs instance. (Tech Lead)
• test/altimate/altimate-base-registration-telemetry.test.ts never calls Telemetry.shutdown() in afterEach, unlike test/telemetry/telemetry.test.ts; leaves the singleton initialised for later files. (Tech Lead)
• session_start is now an anchor event, so a serve process hosting N sessions issues up to N immediate flushes instead of batching; watch App Insights ingestion volume. (CTO)
• ALTIMATE_CLI_COMMAND is written to process.env and inherited by every spawned MCP server / shell subprocess. Value is a closed enum, no exposure, but it is new unscoped env surface. (CTO, Gremlin)
• startLoopMonitor() runs in the TUI worker even when telemetry is disabled (250 ms unref'd timer producing dropped events); harmless but inconsistent with the main thread, which starts it inside doInit(). (End User)
• cold-start-regression CI job is linux-x64 only. (PM)
• needsDependencies does sync fs work inside the config Effect layer: deliberately bounded (one-level glob); add a comment so nobody makes it heavier. (Tech Lead)
• Local test env: test/mcp/headers.test.ts and test/server/httpapi-mcp.test.ts fail on a developer machine whose global config declares MCP servers (the real servers show up in TestInstance's status). The tests should isolate the global config dir. (Release driver)