render.py resolved `vouch` off PATH first, so a stale global install
shadowing the venv would render every shot against the wrong build and
silently overwrite the committed images. prefer the vouch next to the
running interpreter (the venv's editable build); fall back to PATH only
when none exists. the guard test now passes without activating the venv.

docs(examples): surface screenshots in readme, harden render resolution

add docs/tutorials/ — an index plus four diátaxis walkthroughs (build your
first knowledge base, give a coding agent a reviewed memory, share a kb via
bundles, edit a kb in obsidian). every command was run end-to-end against
the cli before it went in, so the walkthroughs only cover verified
behaviour (including the real forbidden_self_approval review gate).

feature the native one-line wire-up `claude mcp add vouch -- vouch serve`
in getting-started step 6, and add a "how a host talks to vouch" section to
docs/transports.md describing the stdio mcp path end to end — spawn,
handshake, tool call, review gate, _meta sidebars, server-to-client push.

fix the read-sidebar guidance in claude.md: the per-tool read sidebar is
_meta.vouch_salience (salience.py, attached by kb_context), not
vouch_hot_memory — that key is written only by the openclaw context
engine's assemble().

docs: add tutorials, host-comms walkthrough, one-line mcp install

`vouch pending` — and every bulk `list_*` path — crashed with a yaml
ReaderError when a single proposal file held a control character: one
unreadable file took down the whole listing. add a `_load_or_skip` helper
that logs a warning and skips the bad file instead of aborting, applied
across proposals, claims, entities, relations, evidence, sessions, sources.

the control character traces back to vouch's own file i/o relying on the
locale default encoding: on a non-utf-8 locale (e.g. latin-1) read_text /
write_text / open mangle non-ascii into raw bytes the yaml loader rejects,
and can even crash on write. pin encoding="utf-8" on all text-mode pathlib
i/o under src/vouch.

adds regression tests for the resilient listing.

fix(storage): survive unreadable artifact files and pin utf-8 i/o

update package version across pyproject.toml, src/vouch/__init__.py,
openclaw.plugin.json, and SPEC.md examples to reflect 1.0.0 stable release.

Chore/version 1.0.0

feat: claude code session auto-capture + approved-knowledge recall

design: automatic finalization of old buffers on sessionstart,
and current session on window close.

plan: 8-task breakdown covering capture.py functions, cli command,
adapter hooks, comprehensive unit tests, integration test, and
documentation.

implement is_stale_buffer() to check file age, and finalize_all_except()
to bulk-finalize capture buffers older than a threshold. this enables
sessionstart cleanup of orphaned buffers from previous sessions.

includes 9 unit tests covering single/multiple buffers, age boundaries,
and current session exclusion.

new command finalizes all capture buffers except the given session,
if they are older than max_age_seconds (default 3600s). silently
succeeds if no kb found. used by sessionstart hook to clean up
orphaned buffers from previous sessions.

includes 3 unit tests covering option parsing, env fallback, and
graceful degradation.

run 'vouch capture finalize-all' as the first sessionstart hook
to clean up orphaned buffers from previous sessions before
banner and recall commands. ensures old sessions are finalized
automatically on next session start.

add comment to sessionstart hook explaining that current session
finalization happens on next session start (fallback behavior when
windowclose event is not supported by the vs code extension).

add section describing how capture works, configuration options,
and the fallback behavior if windowclose event is not available.

…rigin

when finalize_all_except() cleans up orphaned buffers, those old sessions
have unknown origin (no cwd), so we should not attribute the current
working directory's git diff to them. this prevents data corruption where
old sessions' summaries get current repo changes, and avoids spurious
proposals that cross min_observations due to misattributed file counts.

added explicit cwd check: if cwd is None (unknown origin), skip git
changes. normal finalize calls with explicit cwd still get git context.

also fix typo: unfinalzed -> unfinalized in adapter README.

feat(capture): automatic session finalization on sessionstart and close

adds kb.detect_themes (read-only cluster detector) and kb.propose_theme
(routes synthesis pages through the review gate). scoring is deterministic
— entity pair co-occurrence across sessions, weighted by log(1 + claim_count).

registered on all four surfaces (mcp, jsonl, cli, capabilities). the cli
exposes `vouch detect-themes` with --propose for one-shot propose-all.

adds THEME to the PageType enum so theme pages pass the page-kind gate.
deduplication prevents re-proposing clusters that already have pending or
approved theme pages.

closes #311

…utput

- config reader now type-checks every value and falls back to defaults
  on malformed input (mirrors salience.reflex_cfg pattern)
- dedup in detect_themes now compares on the resolvable entity subset
  so it stays consistent with what propose_theme actually stores
- --propose --json no longer interleaves human-readable lines on stdout
- adds changelog entry under [unreleased]

feat(themes): cross-session pattern detection via entity co-occurrence

the benchmark suite was implemented but unrunnable out of the box:
pytest-benchmark wasn't in the [dev] extras, and the bench_*.py
filenames don't match pytest's default python_files glob, so the
documented invocation collected zero tests.

- add pytest-benchmark to the [dev] optional-dependencies
- add a `make bench` target that folds in the python_files override
- record the first baseline (medians at 1k/10k) in benchmarks/README.md
  and flip its status from "not implemented yet" to implemented
- gitignore the bench.json / .benchmarks run artifacts

# Conflicts:
#	Makefile

chore(bench): wire make bench, add pytest-benchmark, record baseline

lead the readme with session auto-capture — the feature that stops each
new agent session from re-doing the last one's work.

- banner.svg: add a "sessions capture themselves" on-ramp that funnels
  into the review gate; swap the approve/reject text labels for icons;
  refresh the example date and kb panel; add a self-capturing pill.
- gittensor.md + readme gittensor section: reframe from "cited record of
  the subnet's scoring rules" to a real miner workflow — capture what a
  session learns about a target repo, approve it, recall it next session
  so the investigation happens once. grounded in the real sn74 loop
  (merged prs, repo allocation, a shuffling whitelist).
- demo.gif/tape + example-session.md: the auto-capture walkthrough.
- .claude/settings.json + commands: reference capture hooks and the four
  vouch slash commands. settings.local.json intentionally omitted.

…miner

docs: redraw hero banner + reframe gittensor around the miner workflow