diff --git a/ROADMAP.md b/ROADMAP.md
index a62c742..70f17bb 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -2,11 +2,13 @@
 
 # Roadmap
 
-**Current:** v0.8.0 (Phase 3 complete)
+**Current:** v0.8.8 shipped (Phase 3 complete; hardening rounds underway, see
+CHANGELOG). Cross-platform capture is the next feature phase.
 
 Screencast MCP ships in phases. Phase 1 is the capture / watch / minimal-edit
 core. Phase 2 adds the full edit surface, safety redaction, and system audio.
-Phase 3 adds the produce / trailer layer. Cross-platform capture is next.
+Phase 3 adds the produce / trailer layer. Phases 1 to 3 are shipped; the work
+since has been correctness hardening. Phase 4 onward is proposed below.
 
 ## Phase 1 -- Capture, watch, minimal edit (shipped)
 
@@ -21,7 +23,7 @@ Phase 3 adds the produce / trailer layer. Cross-platform capture is next.
 - [x] Quality presets (draft / standard / high) instead of raw ffmpeg flags
 - [x] Threat-model documentation
 
-## Phase 2 -- Full edit surface and audio
+## Phase 2 -- Full edit surface and audio (shipped)
 
 - [x] `crop`, `scale`, `speed`, `overlay`, `compress`
 - [x] `extract_audio` (extract the audio track) and `clip` (frame-accurate
@@ -32,13 +34,76 @@ Phase 3 adds the produce / trailer layer. Cross-platform capture is next.
 - [x] System audio capture (dshow loopback) via the `start_recording` audio
       option and `list_audio_devices`. Microphone capture is deferred indefinitely.
 
-## Phase 3 -- Produce / trailer layer
+## Phase 3 -- Produce / trailer layer (shipped)
 
 - [x] `assemble_highlights` (stitch clips with cuts or transitions) and
       `xfade_transition` (crossfade two clips), with auto-normalized inputs
 - [x] `title_card` (bundled font) and `music_bed` (loop, fade, mix, duck)
 - [x] Aspect variants (`reframe`) and platform export presets (`export_preset`)
 
-## Cross-platform
+## Hardening -- correctness rounds (ongoing, v0.8.1+)
 
-- [ ] Capture backends beyond gdigrab (avfoundation on macOS, x11grab on Linux)
+Bug-fix and robustness work on top of Phase 3, one focused PR per issue. Details
+in CHANGELOG; representative items:
+
+- [x] Recording lifecycle: a second server instance no longer kills another's
+      live recording (#16); concurrency-aware, bounded session registry (#17)
+- [x] Produce edge cases: `assemble_highlights` clip-duration vs transition
+      validation (#18) and audio preservation on mixed inputs (#19)
+- [x] `sample_frames` returns only frames this run actually wrote, in both fps
+      (#20) and timestamps (#35) mode
+- [x] Capture geometry: DPI-aware monitor enumeration (#39), odd region
+      dimensions rounded to even for recording (#34), off-desktop region offsets
+      rejected with a clear error (#41)
+- [x] Audio device enumeration is capped and gives an actionable error instead
+      of a 30s hang (#40)
+- [x] Filter-string validation up front (#22); temp working files written to the
+      OS temp dir (#24); ESLint + Windows CI + typecheck in CI (#21, #23)
+- [ ] Open: odd output dimensions in edit/produce tools (#47), `concat` relative
+      input paths (#48), `extract_audio` `copy` container mismatch (#36),
+      overwrite protection for explicit output paths (#37)
+
+## Phase 4 -- Cross-platform capture (next)
+
+Capture today is Windows-only (gdigrab). Generalize the capture backend behind
+the existing `Provider` interface so the same tools work on macOS and Linux.
+
+- [ ] macOS capture via avfoundation (screen + window + region)
+- [ ] Linux capture via x11grab, plus a PipeWire/Wayland path where X11 is absent
+- [ ] Per-platform monitor/window enumeration behind one resolver (today's
+      PowerShell shims are the Windows implementation of that seam)
+- [ ] Capability reporting so a tool returns a clear "not supported on this
+      platform" instead of a cryptic backend failure
+- [ ] CI capture smoke tests on the macOS and Linux runners
+
+## Phase 5 -- Captions and annotation (proposed)
+
+Screencast-native polish that turns a raw capture into something watchable.
+
+- [ ] `burn_subtitles` -- burn an SRT/ASS track into the video (styleable)
+- [ ] `transcribe` -- optional speech-to-text to an SRT sidecar (pluggable
+      engine; opt-in, no bundled model), feeding `burn_subtitles`
+- [ ] `zoom_to` / cursor emphasis -- animated zoom-and-pan to a region, and a
+      cursor highlight/spotlight, the highest-value screencast affordance
+- [ ] `callout` -- timed text/arrow/box annotations over the video
+- [ ] `keystroke_overlay` -- render captured keystrokes/clicks as on-screen badges
+      (requires a capture-time input hook; scope TBD)
+
+## Phase 6 -- Audio polish and delivery (proposed)
+
+- [ ] `normalize_audio` -- EBU R128 loudness normalization (`loudnorm`)
+- [ ] `denoise` -- background-noise reduction for narration
+- [ ] `thumbnail` / `contact_sheet` -- single poster frame or a grid of frames
+- [ ] Optimized GIF/WebP export (palette tuning, size budget) beyond `convert`
+- [ ] Chapter markers / cue points written into the container
+- [ ] `pipeline` -- chain several edits in one call to avoid intermediate files
+
+## Cross-cutting backlog
+
+Not tied to a single phase; pull in opportunistically.
+
+- [ ] Overwrite protection: refuse to clobber an existing explicit `output`
+      unless `overwrite: true` (#37)
+- [ ] Node `engines` already declared; document the supported range in
+      CONTRIBUTING
+- [ ] Consistent input-path resolution (absolute) across all tools (see #48)