Update e-dant/watcher 0.13.8 -> 0.14.5

[rsanheim]

Bumps the embedded e-dant/watcher from 0.13.8 to 0.14.5 and re-vendors all four platform binaries.

Why

0.14.5 is event-correctness / de-duplication work over 0.13.8:

• macOS FSEvents: create-event de-dup no longer interferes with other event types (fix(minitest): tag test-written stdout instead of guessing #106)
• Linux inotify: rename-association buffer + partial rename support, duplicate-create fix, cross-directory path caching fix (fix(minitest): stop discarding stdout written by the tests themselves #105, Prepare changelog for v080-rc1 #108, CircleCI: parallelize test-all and add Go/Ruby caching #109)
• Fixes a regression that 0.13.8 itself introduced (upstream fix(minitest): stop discarding stdout written by the tests themselves #105 via Explicit color options: --color=auto|always|never #89), so this is strictly more correct than what we shipped

Net effect for plur: fewer or equal file events, never a new field or event type.

What changed

• lib/plur/config.rb — edant_watcher_version 0.13.8 -> 0.14.5 (single source of truth)
• embedded/watcher/* — all 4 platform binaries re-vendored via rake vendor:clean + rake vendor:download:all (checksums verified on download)
• docs/features/watch-mode.md — document the create/modify effect-type filter and why a bare mtime touch does not trigger a run

No parser or test changes: the CLI invocation and stdout JSON event format (path_type, path_name, effect_type, associated) are unchanged between the two versions, and the doctor spec already masks the version line.

Verification

• macOS (arm64): plur doctor reports 0.14.5; bin/rake full suite green (0 failures); raw binary emits the expected JSON.
• Linux (arm64, via Docker/Ubuntu): fresh build embeds watcher-aarch64-unknown-linux-gnu, doctor reports 0.14.5, rspec/minitest/mixed fixtures + a real external gem pass, and plur watch re-runs on a content edit (and correctly ignores a bare touch).

Draft pending cross-platform CI (the Linux/Windows binaries can't be smoke-tested locally beyond the arm64 Docker run).

[rsanheim]

Upstream changelog: 0.13.8 → 0.14.5

Full upstream entries between our current pinned version (0.13.8, on main) and the new one (0.14.5), grouped by whether they affect plur. plur ships the standalone watcher CLI binary on macOS/Linux and parses its stdout — so the file-event-correctness fixes matter to us, while the Go-package / Windows / CMake / nodejs items do not change how we consume it.

Directly relevant — file-event correctness (macOS FSEvents + Linux inotify):

• 0.14.5 — inotify (Linux): fixed cross-directory path-name caching/reporting (CircleCI: parallelize test-all and add Go/Ruby caching #109).
• 0.14.4 — inotify: fixed rename reporting when the "moved-from" half is missing (fix(minitest): stop discarding stdout written by the tests themselves #105, a regression introduced by 0.13.8's Explicit color options: --color=auto|always|never #89 fix) + regression test; fsevents (macOS): fixed a rare create-event duplication when a rename immediately follows a create on the same path; Windows rename path + IO-loop fixes.
• 0.14.3 — inotify: fixed duplicate creation events when an event matched both the directory-creation path and the non-pending parse result.
• 0.14.2 — inotify: fixed rename reporting when the associated event is still pending (fix(minitest): stop discarding stdout written by the tests themselves #105); fsevents: prevented create-event de-duplication from interfering with other event types (fix(minitest): tag test-written stdout instead of guessing #106).

Not relevant to plur's usage (we don't use the Go bindings, the library-as-submodule, or Windows DLL exports):

• 0.14.1 — Watcher-Go module import fix (build(deps): bump websocket-driver from 0.8.1 to 0.8.2 in /fixtures/projects/default-rails #101).
• 0.14.0 — new Go package (fix(rspec): renumber aggregate-failure sub-markers instead of leaking ‽ #96/build(deps): bump loofah from 2.25.1 to 2.25.2 in /fixtures/projects/default-rails #100), watcher-c DLL exports on Windows (build(deps): bump websocket-driver from 0.8.0 to 0.8.1 in /fixtures/projects/default-rails #92), CMake "good neighbor when embedded" changes (Prepare version 0.72.0 #93). Also Windows dist archive header layout (Prepare changelog for v080-rc1 #108, in 0.14.4).
• CMake minimum bumped 3.9 → 3.11; nodejs project dev-dependency bumps.

Net for plur: the CLI invocation and stdout JSON event format (path_type, path_name, effect_type, associated) are unchanged across the whole range, so no parser or test changes are needed. The behavioral delta is strictly fewer/cleaner events (de-dup + rename-association fixes), and 0.14.5 undoes a regression 0.13.8 itself shipped (#105 via #89).

Manual verification

Linux (arm64, Ubuntu 24.04 via Docker): fresh go build embeds watcher-aarch64-unknown-linux-gnu; plur doctor → 0.14.5; rspec/minitest/mixed fixtures + a real external gem all pass; plur watch re-runs on a content edit and ignores a bare touch. Raw inotify JSON matches the expected schema.

macOS (arm64): fresh build from this branch (402182f3):

• plur doctor → File Watcher 0.14.5, darwin/arm64.
• Real third-party repo (janlelis/paint) cloned + bundled; baseline plur -n 4 → 56 examples, 0 failures.
• Live plur watch session against that repo, with each action timestamped in a separate log and correlated against plur's debug events:
  • touch (settled file), twice → no event, no rerun
  • content edit → modify event → spec re-ran (17 examples, 0 failures)
• Note on method: an initial un-settled run looked like a touch triggered a rerun, but that was FSEvents delivering the backlog of git clone create events ~5s into the session — not the touch. Re-running with an 8s settle and separate op-timestamps confirmed touch does not trigger a run. This is the create/modify-only effect-type filter in cmd_watch.go doing its job.

Both platforms confirm the upgrade is drop-in: correct version reported, tests run, watch mode behaves identically to before (only with the upstream event-correctness improvements underneath).