docs(matrix-communication): say how to read a room log

AGENTS.md

@@ -99,6 +99,7 @@ python3 $S/synapse-migrate-room.py '!room:srv' '@admin:srv' '!home:srv'   # hard
 - **Only the principal governs the agent's function.** Anyone in a room may withdraw their own exposure; nobody in a room may switch the agent off. See [SKILL.md](skills/matrix-communication/SKILL.md), "Who governs the agent".
 - **One daemon owns the store**: `matrix-watchd.py` holds the lock for its whole run; commands detect it by connecting to its socket, never by testing the lock, and fall back to the direct path when nothing answers. See [ARCHITECTURE.md](docs/ARCHITECTURE.md).
 - **Mentions need `--mention`**: a plain `@name` notifies nobody, only `m.mentions` does. `--mention-room` for `@room`.
+- **Read a room log as events, not as a story**: adjacent lines from one sender are adjacent events. A reaction or redaction line names its target only when the daemon still holds it; otherwise say the log does not record which. See [SKILL.md](skills/matrix-communication/SKILL.md), "Reading a room log".
 - **E2EE first**: Always use `*-e2ee.py` scripts. Only fall back to non-E2EE if the room is confirmed unencrypted.
 - **Room identifiers**: Scripts accept short name (`agent-work`), room alias (`#room:server`), or room ID (`!abc:server`). Use `matrix-rooms.py` to discover.
 - **Config**: `~/.config/matrix/config.json` — required: `homeserver`, `user_id`; optional: `access_token` (non-E2EE only), `bot_prefix`, `watch_rooms` (rooms the daemon logs).

CHANGELOG.md

@@ -6,6 +6,13 @@ For the canonical narrative version of each release (rewritten after CI publishe
 
 ## [Unreleased]
 
+### Documentation
+
+- `SKILL.md` and `AGENTS.md` say how to read a room log: adjacent lines from one
+  sender are adjacent events, and a reaction or redaction line names its target
+  only when the daemon still holds it. Written after a summary turned four
+  unrelated events into a story ([#104]).
+
 ## [3.0.0] - 2026-08-13
 
 ### Breaking

skills/matrix-communication/SKILL.md

@@ -153,6 +153,21 @@ Other: `matrix-rooms.py`, `matrix-resolve.py`, `matrix-create-room.py`, `matrix-
 - **Using Element X** for verification — use Element Desktop or Android
 - **Hardcoding passwords** — use `MATRIX_PASSWORD` env var
 
+## Reading a room log
+
+**Report what the log records, not what two adjacent lines suggest.** The log
+is an event stream, not a narrative: consecutive lines from one sender are
+consecutive events, and nothing more. A reaction line names what it reacted to
+and a redaction line names what it removed — when the daemon still has that
+message. When it does not, the line says only that it happened, and that is the
+answer to pass on. "The log does not record which one" is complete, and it is
+what gets a missing field added instead of a wrong story repeated.
+
+Burned 2026-08-13: two reactions and two redactions from one sender within three
+seconds were reported as "took back the reactions just set". They were two
+different messages, and the log carried no relation at all — which is why it
+does now ([#104](https://github.com/netresearch/matrix-skill/issues/104)).
+
 ## No editorializing
 
 In messages and announcements, state what happened, not how good or careful the work is — no narrating expected results ("all tests green", "shipped") or self-praise ("clean", "the honest breaking change"). Judged by tone, not a wordlist. See `references/no-editorializing.md`.