chore(backup): adopt backup/sessions/ convention

.gitignore

@@ -199,3 +199,11 @@ ENV/
 # GNAT Pretty Print temp files
 *__GNATPP-TEMP
 *.tgz
+# Claude Code session exports / ephemeral scratch space.
+# The CLI /export command has been broken for some time (writes
+# 0-byte stubs); durable session artifacts now live in
+# backup/sessions/ via the session_snapshot tool in
+# scripts/python/shared. The exports/ directory is kept as a local
+# scratch path for cross-project context bouncing, but its contents
+# are not git-tracked. See backup/sessions/README.md.
+exports/

backup/sessions/README.md

@@ -0,0 +1,50 @@
+# backup/sessions/
+
+Durable, git-tracked archive of session artifacts for this project.
+
+## What lives here
+
+Three kinds of files, all using the same filename convention so
+chronological sort works across kinds:
+
+```
+<YYYYMMDDTHHMMSSZ>__<original-basename>
+```
+
+| Kind | How it gets here | Source of the content |
+|---|---|---|
+| `*_alignment_note.md` | Written directly by the assistant when a slice opens with a design decision. Committed with the slice's first commit. | The assistant's pre-coding review of the frozen docs + GPT check. |
+| `*_pr_review_package.md` | Written directly by the assistant before a PR is opened. Committed with the slice's PR. | The assistant's self-contained review document for GPT. |
+| `*` (time-driven snapshots) | Mirrored here by `scripts/python/shared/session_snapshot`, typically via the `/snapshot` custom slash command. | A Claude Code memory file from `~/.claude/projects/<encoded-project>/memory/`, preserved verbatim after the `<timestamp>__` prefix. |
+
+Filenames sort naturally by UTC timestamp regardless of how many
+different kinds coexist in the directory. Recovery of any file is a
+plain `cp` with prefix strip (see
+`scripts/python/shared/session_snapshot/README.md` for the
+`--restore` tool or the manual procedure).
+
+## Relationship to `backup/sessions/raw/`
+
+`raw/` is **gitignored** and holds compressed full-session `.jsonl`
+backups produced by `scripts/python/shared/jsonl_snapshot`. Those
+files are large (~10 MB per snapshot) and intended for external
+storage. Git-tracked files stay up here in `backup/sessions/`.
+
+See `scripts/python/shared/jsonl_snapshot/README.md` for the full
+forensic-tier workflow.
+
+## What this is NOT
+
+- **Not the Claude Code `/export` command output.** The CLI
+  `/export` has been broken for some time; this directory is the
+  workflow that replaces it.
+- **Not a full conversation transcript.** Memory files are
+  curated recap summaries written by the assistant at key
+  workflow points. Use `jsonl_snapshot` and the raw session
+  `.jsonl` under `~/.claude/projects/.../<uuid>.jsonl` for
+  forensic recovery of verbatim conversation content.
+- **Not a replacement for the Claude Code memory system.**
+  Memory files in `~/.claude/projects/.../memory/` remain the
+  auto-loaded source of strategic context at session start.
+  This directory is the off-machine-durable backup copy that
+  survives a local-state loss.

backup/sessions/raw/.gitignore

@@ -0,0 +1,12 @@
+# Everything in this directory is gitignored on purpose: compressed
+# session jsonls are large binary blobs that belong on external
+# storage, not in git history. The directory itself is tracked
+# (via this .gitignore file being the only tracked file) so fresh
+# clones get a usable target for jsonl_snapshot without needing to
+# run mkdir.
+#
+# See ../README.md and scripts/python/shared/jsonl_snapshot/README.md
+# for the workflow that targets this directory.
+
+*
+!.gitignore