v0.33.0

Changed

• Changed: the SessionStart hook does less before it answers. Work whose result
  the first turn does not need -- previous-session recovery, the capture-gap
  check, the slug and case-divergence records, publishing the start-context
  cache, and the before_session_start dispatch when git_restore.enabled is
  false and no third-party hook is registered -- now runs in a detached phase
  after the memory context has been emitted. Rendering a memory file under 16KB
  no longer forks cat, REMEMBER_ROOT no longer forks dirname, and
  mkdir -p on an existing tmp/ is skipped. Foreground subprocess count per
  start drops from 49 to 29; on Git Bash, where a fork costs 10-50x what it
  costs on Linux, that is most of the hook's own cost. Measured on CI: the
  hook's traced cold run on windows-latest falls from 2.32s to 1.60s on a
  baseline store and 2.87s to 2.16s on a 20MB one; on ubuntu-latest from 148ms
  to 102ms. Part of #660.
• Changed: REMEMBER_DEFER=0 runs that deferred phase inline instead. It
  exists for tests that assert on one of those side effects and need it to have
  happened by the time the hook exits; the default is deferred. Part of #660.
• Fixed: two concurrent dispatch calls shared one capture file
  (tmp/dispatch-stdout.$$ -- $$ is the shell's pid and does not change in a
  subshell), so the background dispatch's cleanup could delete the file the
  foreground one was still writing and a plugin's injected context vanished
  with no error on any channel. The capture files are now named per event.
  Found while deferring the before_session_start dispatch for #660.

Fixed

• Fixed the Windows Git-Bash benchmark's hook-dir probe: it invoked the literal bash (which resolves to the WSL launcher and emits UTF-16LE/NUL-interleaved output, exit 127 on a missing distribution) instead of resolve_bash()'s Git Bash, and its output decoding is now UTF-16LE-aware (#672).

• Fixed: the handoff_mode: "per_session" degrade notice (#363) silently
  stopped printing in default (in-project) storage mode when no session_id
  reached the SessionStart hook. The notice fired only in external storage
  mode or when a session_id resolved, so a user on default storage whose
  per_session request quietly fell back to the shared remember.md got no
  indication of it. Fixes #686.

• Fixed: under a Turkish locale (tr_TR, and az_AZ the same way) no save ever completed. safe_eval in scripts/log.sh matches the pipeline's KEY=VALUE output with ^([A-Z_][A-Z0-9_]*)=, and [A-Z] is a POSIX bracket range matched by the locale's collation rather than by byte value -- Turkish collation does not place I inside A..Z, so every bridge variable whose name carries an I (EXTRACT_FILE, POSITION, SKIP_LINES) was silently dropped while the counts still arrived. The run passed its "0 exchanges" gate and died in build-prompt on FileNotFoundError: '', thousands of times over. safe_eval now matches under LC_ALL=C for the duration of the function, the same guard config() already carries in that file (#695).

• Fixed: scripts/save-session.sh now stops at the bridge, with an error naming EXTRACT_FILE, instead of calling build-prompt with an empty path. Whatever the reason the extract step's output loses a variable, the report should name the crossing rather than surface as a Python traceback from open('') (#695).

• Fixed: a memory file whose size was never measured was emitted through bash's read rather than cat. _remember_wc_size_get_into returned 0 for a path it held no measurement for, and the render loop rewrote anything non-numeric to 0 as well, so _remember_emit_file's own "no usable size" arm was unreachable and the 16 KB threshold added in this same release was defeated in exactly the degraded case it was written for -- a 4 MB file on the session-start foreground path, measured at 24.1s on windows-latest. The batched wc -c can fail wholesale, which leaves every file unmeasured at once. An unmeasured size is now kept distinct from zero at all three of the getter's call sites, and each says size unknown rather than printing 0 bytes -- or an empty ( bytes) -- for a file nobody measured (#695).

• Fixed: the #695 source scanner credited a nested function's local LC_ALL=C to its enclosing function. _function_spans closed a function at the first } at column 0, and #660's deferred-phase work defines capture_was_seen at column 0 inside _remember_deferred_phase -- so 148 lines that hold no such declaration were reported as protected. Neither change contained the defect on its own; it existed only once both had landed. The parser now tracks nesting with a stack and attributes a declaration only to the lines it governs (#695).

• Fixed: the capture-gap notice's own tests raced the phase that writes it. #660 moved the check into the deferred phase, where it is the last thing that phase does, so a hook that has returned says nothing about whether the check has run -- the positive assertions failed intermittently (caught on macos-latest 3.10, green on the other eleven legs) and, worse, all six negative assertions passed just as well against a phase that had not reached the check yet or had died before it. The positives now poll on the default deferred path, which is what proves the phase reaches its end; the negatives run the phase inline, where "did not fire" is a verdict rather than a race (#695).

• Fixed: a config key containing an I disabled the config-data cache outright on a Turkish-locale host. _remember_cfg_flatten_cache_valid_line validated each cache line with [[ =~ ^_RCFG_[A-Za-z0-9_]+= ]], and inside [[ =~ ]] a bracket range is matched by the locale's collation rather than by byte value -- Turkish collation does not place I inside A..Z. The line read as malformed, the cache was refused, and every hook fell back to re-reading and re-flattening the config on every invocation, permanently, with nothing said (#695).

• Fixed: _remember_normalize_win_path stopped recognising the Windows drive I: under the same collation, and its tr '[:lower:]' '[:upper:]' mapped a lower-case i: drive to a dotted İ under Turkish case rules -- a two-byte character where one ASCII letter has to go. The tr needed its own LC_ALL=C prefix rather than the function's local: local on a name the environment never exported leaves it unexported, so on a host whose language is set through LANG alone the child process kept the user's locale (#695).

• Changed: eleven further bracket ranges across scripts/ and hooks.d/ now match under LC_ALL=C, scoped to their own function. Measured on glibc, only [[ =~ ]] collates -- case patterns and ${v//[!...]/} compare bytes -- so these were not broken, but one line makes them immune and the rule is now uniform. A new source scanner (tests/test_locale_ranges_695.py) fails the suite when a letter range appears outside an LC_ALL=C scope, with a written reason required for each allowlisted exemption (#695).