feat(agent): a missing workspace gates the file tools, not the whole run - #78

Follow-up to #76, which fixed the two instances in the code it introduced and left the
older ones alone to keep that review diff readable.

`executeCommand` returns a Thenable, so a bare call in a void context turns any rejection
into an unhandled promise rejection in the extension host — attributed to nothing, which
is the part that makes it useless.

SIX SITES, NOT SEVEN. The brief for this change listed the `levelcode.ai.focus`
registration as needing a `return`. It does not: `() => vscode.commands.executeCommand(…)`
is a concise arrow body, so it already returns the thenable and VS Code already reports its
failures. Left alone.

The other six are all BACKGROUND reveals — addSelection, addContext, resumeSession, the
editor panel's onDidDispose, the sessions view's newSession, and handleLaunch. In each the
accompanying work has already succeeded by the time the reveal runs, so failing that work
because the panel would not come forward would be worse than the panel not coming forward.
They route through one `focusChatView(why)` helper that logs and never rejects.

Logged, not swallowed: the complaint was "attributed to nothing", so `.catch(() => {})`
would answer the letter of it and none of the substance. `why` names the caller in the log.

ONE GUARD FOR THE WHOLE CLASS. Rather than six assertions naming six functions, the test
scans every occurrence of the call and accepts only returned / awaited / concise-arrow /
helper-wrapped / inline-handled forms. That covers the seventh site and the eighth nobody
has written yet — which matters, because this pattern reached six copies precisely because
nothing was watching for it.

Verified non-vacuous against develop: the same scan reports exactly the six bare sites
there (lines 423, 468, 1055, 2298, 2435, 2531) and zero here.

Bypasses, each reverted and confirmed to fail:
  - a single site returned to a bare call; the switch-case site returned to a bare call
  - the focus command wrapped in a block body so it stops returning (proves the guard
    protects the site that was already correct)
  - the helper swallowing silently; the log dropping its caller attribution

One assertion deliberately relaxed after bypassing it: pinning `.then(undefined, …)` over
`.catch(…)` failed a refactor with no behavioural difference. It now accepts either and
still fails when the handler stops logging.

23 tests in chatSurface, 32 suites green.

…lly sent

Both review points on #78 were right.

1. THE TOKEN ESTIMATE STILL COUNTED THE FULL TOOL LIST

`toolsTokensEst` falls back to a module constant on the plain path (no MCP, no recall) so
that path never re-stringifies. But that constant was built from TOOLS, and after this PR
there are TWO plain paths — a rootless run sends PORTABLE_TOOLS and was billed for TOOLS.

Measured: ~1000 tokens, about two thirds of the tool budget, reported against a window
that never spent it. That is worse than a missing feature; it is a meter reading high, and
the context popover exists precisely so that number can be trusted.

This is the same mistake as leaving baseTools on TOOLS, one line further down — which I
fixed, described in the PR body, and then missed here. PORTABLE_TOOLS_TOKENS_EST is a
second constant rather than a call-time derivation, because keeping the plain path free of
JSON.stringify is the whole reason the constant exists.

2. THE TEST DID NOT PIN read_command_output

It asserted seven of the eight gated tools. read_command_output was the omission, and it is
the plausible one to lose: it takes no path and reads as portable at a glance, so nothing
would have objected to un-gating it. Rootless it can only ever refer to a background
run_command that could not have started.

Guards, each bypass-verified by reverting the fix:
  - read_command_output un-gated (the second review point)
  - the estimate reverting to the full-TOOLS constant (the first)
  - the estimate no longer switching on the root; the rootless estimate deleted
  - PORTABLE_TOOLS no longer a filtered subset, which would make the estimate guard vacuous

6 tests in agentNoWorkspace, 33 suites green.