Promote the Unreleased section to 1.2.0 and capture the UX /
operability work landed since v1.1.0:

- #20 install/update progress streaming with -v
- #21 gRPC server wire-up in helper
- #22, #26 test coverage (cli/output 0→82%, systray 0.2→3.2%)
- #24 RefreshOnStart field removal
- #27 bubbles v0.21 -> v1.0
- #28 go:embed baseline catalog (fresh `go install` works offline)
- #29 pkg/logging slog wrapper wired into main and grpc recovery

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix: address Copilot review feedback — correctness fixes

Seven real-bug fixes called out across review comments on merged PRs:

- pkg/platform/platform.go: resetLookPathCache() no longer reassigns
  the package-level sync.Map (unsafe vs concurrent readers). Uses
  Range+Delete instead.
- pkg/installer/providers/brew.go: GetLatestVersion no longer caches
  context-canceled / deadline-exceeded errors. A caller with a short
  ctx would have poisoned the 5m TTL for every later caller.
- pkg/installer/providers/progress.go: the progress writer is now
  wrapped in a locking writer so concurrent Write calls from
  cmd.Stdout and cmd.Stderr (os/exec writes those from separate
  goroutines) don't interleave bytes mid-call.
- pkg/catalog/embed.go: EmbeddedJSON() now returns a copy of the
  backing bytes. Previously callers could mutate the package-level
  []byte, corrupting the baseline for every future Get().
- pkg/catalog/manager.go: loadEmbedded now also probes
  /usr/share/agentmgr/catalog.json — where the goreleaser nfpm
  config installs the catalog for .deb/.rpm users. Previously that
  path was silently ignored at runtime even though goreleaser wrote
  the file there at install time.
- internal/cli/doctor.go: Catalog Sources section updated with the
  matching /usr/share/agentmgr row so `doctor` output mirrors the
  loader's probe order.
- internal/cli/agent.go: removed the --all (-a) and --check-updates
  flags from `agent list`. Both bound to local vars that were never
  read anywhere, so the flags did nothing; users setting them got a
  silent no-op instead of a useful behavior.

Supersedes the corresponding Copilot annotations on PRs #19, #28,
#29, #33, plus the dead-flag note on the agent list command.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore: address remaining Copilot review nits (docs + tests)

Batch of small review-comment fixes — documentation accuracy, test
hardening, and gosec G104 on grpc teardown:

- pkg/api/grpc/server.go: explicit \`_ = s.listener.Close()\` in Stop()
  and ForceStop() with a comment noting the close error is not
  actionable during teardown (silences gosec G104).
- pkg/catalog/manager.go: Refresh doc now explicitly notes that
  singleflight does not merge contexts and the first caller's ctx
  governs the shared fetch — a caller with a short deadline can
  cancel the in-flight work for all shared callers.
- pkg/installer/providers/progress.go: WithProgressWriter doc clarifies
  why passing nil is a no-op (asymmetry lets middleware attach a writer
  safely without downstream code accidentally clobbering it).
- internal/systray/cli_uninstall_darwin.go: file is darwin-only via
  //go:build tag, so platform.ID() is always Darwin at runtime. Removed
  the dead Linux and Windows branches (the filepath/os/platform imports
  they needed are gone too).
- CHANGELOG.md: "(Removed in [Unreleased].)" was a broken shortcut
  reference link — promoted to "(Removed in v1.2.0.)".
- README.md: feature bullet uses the full "agentmgr agent list" command
  prefix, consistent with everywhere else in the doc.
- internal/cli/output/spinner_test.go: non-TTY test replaced the
  30ms sleep with a direct assertion on s.isTTY. Start() is a no-op
  when isTTY is false, so sleeping was measuring nothing.
- internal/cli/output/table_test.go: ANSI-alignment loop now actually
  asserts inter-column padding instead of discarding each line.
- internal/systray/handlers_test.go: mustParseVersion panics on parse
  error so test-setup bugs surface at the fixture, not downstream.
  TestDialogTracking uses bare *exec.Cmd values (trackDialog only
  checks pointer identity) to avoid the PATH lookup for \`true\` on
  Windows CI.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* test(cli): drop "all" flag assertion from list; assert removal

The --all / -a and --check-updates flags were removed from
\`agent list\` in the parent commit because both bound to unused
local vars. Drop the existence assertion for "all" and add an
anti-regression check that asserts both flags are ABSENT.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Five catalog entries (block-goose, openclaw, droid, claude-squad,
amazon-q-cli at lines 468/706/2149/2212/2630) declare cask-ness via
metadata.cask: "true" rather than the canonical metadata.type: "cask".
parseBrewPackage only checked the canonical key, so "brew install"
ran without --cask for those agents and Homebrew rejected them as
unknown formulae.

Now accepts either form:
  - "type": "cask"  (canonical)
  - "cask": "true"  (legacy, common in user-customized catalogs too)

Provider behavior change is the entire fix — no catalog edits needed.
brew install/upgrade/uninstall already build their own arg lists
from isCask, so the catalog command/update_cmd strings (which lack
--cask for these legacy entries) are documentation only.

Closes the corresponding Copilot review comment from PR #28.

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>