diff --git a/accessibility/docs/scratch.md b/accessibility/docs/scratch.md
index 070f5a5..ce6941e 100644
--- a/accessibility/docs/scratch.md
+++ b/accessibility/docs/scratch.md
@@ -2,7 +2,20 @@
 
 **Read this first on any handoff.** Living document: update it whenever state changes, before ending a turn.
 
-Last updated: 2026-08-01 · Currently: **Phase 0 complete. Phase 1 next.**
+Last updated: 2026-08-01 · Currently: **Phases 0–1 done, menus + dialogue done.
+Phase 2 (tile examination) next.**
+
+**Game is now playable start-to-finish by ear** from main menu through the intro
+conversation into gameplay — verified by the testbed, 9/9 checks.
+
+### Run the testbed (do this after any change)
+```
+cd "D:/code/Sulphur Memories Alchemist 0.4.2 sanctified"
+rm -rf ~/AppData/Roaming/LOVE/Alchemist        # start from a clean save
+"C:/Program Files/LOVE/lovec.exe" . --a11y-test
+```
+It drives the game by code via `love.event.push`, navigating by *reading* focus
+rather than counting keypresses. Every bug listed below was caught by it.
 
 ---
 
@@ -197,13 +210,41 @@ fold and review keys.
   (bound, immobilized). Walking proves nothing about the console patch; drive
   `console.print` directly to test it.
 
+### Done — dialogue narration + wall tone retrigger fix
+
+**`ingame_dialogue` descriptor.** Speaks speaker + message on a *node* change, but
+only the response on a *selection* change (re-reading the whole speech every time
+you move between responses would be unusable). Responses can also be picked
+directly with number keys 1–0, which the game already supports.
+
+Two more bugs found and fixed:
+
+4. **Nested `states.switch` announced in reverse.** `ingame.open()` sees
+   `world.pendingDialogue` and switches straight to `ingame_dialogue`. Since we
+   announce *after* `realSwitch` returns, the inner announcement ("Conversation")
+   fired first and the outer ("Game") then interrupted and clobbered it.
+   **Fix: after switching, if `states.getCurrent() ~= newState`, stay quiet** —
+   a nested switch already announced.
+
+5. **Double numbering.** The game already prefixes response text with its number
+   (`"1. (Continue)"`, matching the number key that selects it), so adding ours
+   gave `"1. 1. (Continue)"`. Now only added when absent.
+
+**Wall tones — retrigger fix (per Talon: "stop previous tones and play new ones").**
+Tones are **2.33 s, mono, 44.1 kHz**. Steps land far faster than that and up to
+four directions sound at once, so they stacked into a continuous drone. Now each
+direction stops its previous source before starting a new one, so it reads as a
+per-step cue layered over footsteps, and the volume you hear always reflects
+where you're standing now. `playing` is cleared on level change.
+
+*(Mono matters: OpenAL only spatialises mono sources. Stereo would not pan.)*
+
 ### Known gaps (next up)
-- **`ingame_dialogue` announces "Game", not "Conversation", and reads no dialogue
-  text.** It appears to be entered via a path that announces `ingame` first.
-  Needs a descriptor + investigation of how it's switched to. **This is the next
-  blocker for actually playing.**
 - `ingame` itself has no descriptor — that's Phase 2 (tile examination).
-- Wall tones still unverified by ear.
+- Wall tone retrigger fix not yet confirmed by ear.
+- `worldmap`, `journal`, `character`, `research`, `inventory`, `craft`, `refine`,
+  `trade`, `actions_menu`, `alchemy_table`, `keybindings_menu`, `config_menu`,
+  `menu_text`, `ingame_message`, `menu_message` still have no descriptors.
 
 ### Next — Phase 2: tile examination
 Patch `views/levelCursorPopup.set` to narrate the game's existing examine mode,
diff --git a/accessibility/init.lua b/accessibility/init.lua
index 7cda114..844b246 100644
--- a/accessibility/init.lua
+++ b/accessibility/init.lua
@@ -29,6 +29,7 @@ local describedStates = {
 	"load_game",
 	"new_game",
 	"narration",
+	"ingame_dialogue",
 }
 
 local initialized = false
@@ -100,6 +101,16 @@ local function hookStates()
 		-- After the switch, so the state's open() has already set up its focus --
 		-- main_menu, for one, only picks selectedOption inside open().
 		pcall(function()
+			-- Guard against NESTED switches. open() can itself switch state:
+			-- ingame.open() sees world.pendingDialogue and goes straight to
+			-- ingame_dialogue. Because we announce after realSwitch returns, the
+			-- inner announcement ("Conversation") happens first and the outer one
+			-- ("Game") then interrupts and clobbers it. If we are no longer the
+			-- current state, the inner switch already announced -- stay quiet.
+			if states.getCurrent() ~= newState then
+				return
+			end
+
 			local descriptor = menus.loadDescriptor(newState)
 
 			if descriptor then
diff --git a/accessibility/states/ingame_dialogue.lua b/accessibility/states/ingame_dialogue.lua
new file mode 100644
index 0000000..1ad1540
--- /dev/null
+++ b/accessibility/states/ingame_dialogue.lua
@@ -0,0 +1,100 @@
+-- states/ingame_dialogue.lua
+--
+-- Conversations. Two things must be spoken, and they change independently:
+--   * the NPC's line (speaker name + message blocks) -- changes when the node does
+--   * the currently selected response -- changes as you press up/down
+--
+-- So this tracks the node itself and announces the full block on a node change,
+-- but only the option on a selection change. Repeating the whole speech every
+-- time you move between responses would be unusable.
+--
+-- Responses can also be picked directly with the number keys 1-0, which the game
+-- already supports and which is the fastest way to work through a conversation.
+
+local descriptor = {}
+
+local lastNodeKey
+
+local function nodeKeyFor(speakerName, blocks, options)
+	local parts = {tostring(speakerName), tostring(#options)}
+
+	for _, block in ipairs(blocks or {}) do
+		parts[#parts + 1] = tostring(block.text)
+	end
+
+	return table.concat(parts, "|")
+end
+
+descriptor.read = function(stateTable, menus)
+	local keypressed = stateTable.keypressed
+
+	local options = menus.getUpvalue(keypressed, "options")
+	local selected = menus.getUpvalue(keypressed, "selectedOption")
+
+	if type(options) ~= "table" or type(selected) ~= "number" then
+		return nil, nil
+	end
+
+	-- speakerName / messageTextBlocks are not touched by keypressed directly --
+	-- it calls updateText(), so they are upvalues of that function.
+	local updateText = menus.getUpvalue(keypressed, "updateText")
+	local speakerName, blocks
+
+	if type(updateText) == "function" then
+		speakerName = menus.getUpvalue(updateText, "speakerName")
+		blocks = menus.getUpvalue(updateText, "messageTextBlocks")
+	end
+
+	local nodeKey = nodeKeyFor(speakerName, blocks, options)
+
+	local option = options[selected]
+	local optionText = option and option.text or nil
+
+	-- Selection moved within the same node: just the response.
+	if nodeKey == lastNodeKey then
+		if not optionText then
+			return nil, nil
+		end
+		return string.format("%s, %d of %d", optionText, selected, #options),
+			nodeKey .. ":" .. selected
+	end
+
+	-- New node: speaker, what they said, then the options.
+	lastNodeKey = nodeKey
+
+	local parts = {}
+
+	if speakerName and speakerName ~= "" then
+		parts[#parts + 1] = speakerName
+	end
+
+	for _, block in ipairs(blocks or {}) do
+		if block.text then
+			parts[#parts + 1] = block.text
+		end
+	end
+
+	if #options > 0 then
+		parts[#parts + 1] = #options .. (#options == 1 and " response" or " responses")
+		for i, entry in ipairs(options) do
+			if entry.text then
+				-- The game already prefixes response text with its number
+				-- ("1. (Continue)"), matching the number keys that select it.
+				-- Adding our own gave "1. 1. (Continue)".
+				if entry.text:match("^%s*%d+%s*[%.%)]") then
+					parts[#parts + 1] = entry.text
+				else
+					parts[#parts + 1] = i .. ". " .. entry.text
+				end
+			end
+		end
+	end
+
+	if #parts == 0 then
+		return nil, nil
+	end
+
+	return table.concat(parts, ". "), nodeKey .. ":" .. selected
+end
+
+return descriptor
diff --git a/accessibility/testbed.lua b/accessibility/testbed.lua
index 56e1f98..3c04a03 100644
--- a/accessibility/testbed.lua
+++ b/accessibility/testbed.lua
@@ -141,14 +141,27 @@ local function scenario()
 
 	-- Intro sequence: narration / dialogue / message screens vary, so just advance
 	-- until we land in gameplay.
+	local sawDialogue = false
+	local dialogueText
+
 	for i = 1, 20 do
 		if currentState() == "ingame" then
 			break
 		end
+		if currentState() == "ingame_dialogue" and not sawDialogue then
+			sawDialogue = true
+			dialogueText = focusText()
+			-- Exercise response navigation before advancing.
+			press("down", "dialogue: next response")
+			press("up", "dialogue: previous response")
+		end
 		press("return", "advancing intro (" .. currentState() .. ") " .. i)
 		wait(10)
 	end
 
+	check("dialogue was described", sawDialogue and dialogueText ~= nil,
+		"text=" .. tostring(dialogueText))
+
 	check("reached gameplay", currentState() == "ingame", "got " .. currentState())
 
 	if currentState() == "ingame" then
diff --git a/accessibility/walltones.lua b/accessibility/walltones.lua
index c26caa4..da9274a 100644
--- a/accessibility/walltones.lua
+++ b/accessibility/walltones.lua
@@ -70,6 +70,10 @@ local directions = {
 
 local sources = {}
 
+-- The currently-sounding source per direction, so a tone is never retriggered on
+-- top of itself. See playTone.
+local playing = {}
+
 local lastX, lastY, lastLevel
 
 --------------------------------------------------------------------------------
@@ -143,6 +147,19 @@ local function playTone(name, dx, dy, distance)
 		return
 	end
 
+	-- One tone per direction at a time: stop the previous one, start the new one.
+	--
+	-- The tones are 2.33s long and steps land far faster than that, with up to
+	-- four directions sounding at once -- so letting them overlap stacked into a
+	-- continuous drone instead of reading as a per-step cue. Cutting the old tone
+	-- keeps this a layer on top of footsteps, and means the distance/volume you
+	-- hear always reflects where you are standing now.
+	local existing = playing[name]
+	if existing then
+		pcall(existing.stop, existing)
+		playing[name] = nil
+	end
+
 	-- Tagged so the game's own sfx/interface volume sliders apply.
 	local ok, source = pcall(TEsound.play, path, "static", {"sfx", "interface"}, volumeFor(distance))
 	if not ok or not source then
@@ -157,6 +174,8 @@ local function playTone(name, dx, dy, distance)
 			lz -- same height as the listener: maximise horizontal pan
 		)
 	end)
+
+	playing[name] = source
 end
 
 function walltones.playFor(level, x, y)
@@ -208,6 +227,7 @@ function walltones.update()
 	if level ~= lastLevel then
 		lastLevel = level
 		lastX, lastY = x, y
+		playing = {}
 		return
 	end