chameleon 2.38.21

Round-2 real-hook QA of the Stop backstop and injection paths, orchestrated as a
six-agent probe workflow over exclusive real repos with each finding
adversarially verified through the real hook binaries. It targeted round 1's
deferred cross-file gaps, the second-order effects of the _stop_gates
restructure (which came back clean), and framework/edge-case depth round 1 did
not reach. Fifteen fixes; several close gaps in round-1 code.

Fixed

• The cross-file existence advisory was blind to non-root workspaces in a
  monorepo. At the repo-root cwd it loaded one reverse/constant index for the
  git top-level, so removing an exported symbol in a sub-workspace (e.g. a
  monorepo's web app) that other files in that workspace still import was silent
  at turn end, even though the edit-time check fired. Each edited file's break is
  now resolved against its OWN workspace index (the same per-file resolution the
  edit-time check uses), keyed and cached per workspace root.
• The deleted-module advisory (crossfile PATH) rendered the module path
  unsanitized, so a crafted deleted filename carrying ANSI escapes, a newline,
  or a forged [chameleon-untrusted-data:] / [🦎 chameleon:] marker reached the
  Stop context verbatim. The module path is now stripped of control bytes and run
  through the context sanitizer like the symbol and site fields already are; every
  single-line path field in the advisory gets the same treatment so a newline can
  no longer split the line.
• A deleted module's break was lost whenever the turn's first Stop
  short-circuited (the once-per-session idiom-review block prunes the deleted
  file from state before the advisory pipeline runs), which silenced the
  deleted-module advisory on any repo with taught idioms. Deletions are now
  persisted for the session and surfaced exactly once on the Stop that reaches the
  advisory pipeline.
• The reviewer-reply parser let a throwaway array shadow the real findings. A
  [1, 2, 3] or an empty [] appearing in reviewer prose before the findings
  array was returned instead of it — the empty case silently reporting "reviewed,
  no bugs." The parser now prefers the first array that contains objects, falling
  back to a lone object reply, then to any array (a genuine empty [] clean pass),
  so junk before the real payload no longer wins.
• A completed rename that left the old name only in inert Ruby text false-fired
  the cross-file advisory. The presence check blanked quoted strings but not #
  comments, %w/%i/%q/%Q/%() literals, or heredoc bodies, so a stale
  mention in any of those read as a live reference. All are now blanked, keeping
  the #{Const} interpolation carve-out (a real reference).
• The cross-file reference check now blanks strings and comments with a proper
  character scan, closing a class of comment/string edge cases the regex approach
  could not. A comment marker inside a string (a URL's //) must not make the
  comment pass swallow real code, and a quote inside a comment (an apostrophe in
  // don't remove) must not open a string that spans the comment and eats real
  code — no regex ordering handles both, and paired apostrophe-comments could even
  over-report a comment-only mention. A single left-to-right scan that recognizes
  comments before strings (so an in-comment quote never opens a string, and an
  in-string marker never starts a comment), handles TS ////* *//template and
  Python #/triple-quote/f-string, preserves the TS # private-field sigil, and
  respects escapes. This subsumes the round-1 comment-blanking and the multi-line
  block-comment fix.
• The Stop block for a Python-only deletion handed a // ignore hint (a
  syntax error in Python): on a pure-deletion turn the surviving-files list was
  empty, so the hint fell through to the C-style default. The deleted path's
  extension now informs the hint token.
• FastAPI stale-test pairing never fired for the standard tiangolo layout
  (backend/app/… ↔ backend/tests/…): the source-root mirror only swapped a
  LEADING app/src/lib. It now also swaps a mid-path source root, so the
  backend/app/api/routes/x.py → backend/tests/api/routes/test_x.py mapping is
  recognized.
• The Rails co-change applicability check was silently disabled on large
  monorepos: its bounded file walk visited alphabetically-last trees first and
  spent its budget before reaching app//db//config/. The walk now visits the
  co-change-relevant trees first (ranked, app ahead of db) and the budget was
  raised so both the trigger and companion sides are sampled.
• The prose read-path injection denylist now also catches email/mail (and
  forward/share) exfiltration verbs and "disregard the conventions" phrasing,
  closing adjacent bypasses of the round-1 broadening.
• SessionStart now resolves the repo from the payload's cwd rather than the hook
  process cwd, so a harness whose process cwd diverges from the session cwd no
  longer risks injecting the wrong repo's conventions.
• Alembic versions/ migration files (under alembic/ or migrations/) are now
  roled as migrations at bootstrap, so their auto-generated globals no longer
  pollute the generic app archetype's reuse list. (Takes effect on re-bootstrap.)
• An unopenable CHAMELEON_HOOK_ERROR_LOG (a broken symlink, or a path under a
  missing parent dir) made every hook silently skip its Python and no-op — the
  round-1 FIFO guard only caught an existing non-regular file. The guard now also
  falls back to /dev/null for a broken symlink or a missing parent dir.
• A rename inside a file re-exported through a barrel (export * from './x')
  was silent at turn end. The reverse index attributes a star-re-exported name
  to the barrel module, not the origin, so removing it from the origin left the
  barrel's importers broken while the origin's own key had no reverse entry. The
  advisory now finds sibling index.* barrels that re-export an edited origin,
  recomputes the barrel's effective export set (expanding its stars one level),
  and reports the broken importers. It fails safe: any unreadable or nested-star
  source suppresses the finding, so a name still provided by another star of the
  same barrel never false-fires, and an importer updated in the same turn is
  dropped by the live re-check.
• The TS export-name reader missed export type { X } from re-exports and
  mis-read an inline export { type Foo, Bar } clause (dropping Foo, adding a
  spurious type), so a re-exported type read as a broken existence-break on a
  clean file — a cross-file false positive, most visible through the barrel path
  above. The export-clause parser now matches the type modifier and strips the
  inline type specifier prefix, so type-only exports are counted like any other
  export.
• The cross-file reference check now reuses the proven TS lexer, adding
  JSX-text handling. The cross-file reference check blanks literal JSX text
  children so a removed export lingering only as <Tag>Name</Tag> text (which
  renders text, not the variable — unlike {Name}) no longer false-fires. It
  works by iteratively collapsing complete JSX elements, which only exist in
  genuine JSX because a closing </ is unforgeable by real code once strings,
  comments, and templates are blanked; {expr} spans are always kept, so a real
  reference in an expression — even nested between closing tags
  (<A><B>x</B>{Foo}</A>) — still fires, and a comparison x > Foo, division
  a / Foo / b, or generic Array<Foo> (none of which forms a complete element)
  is never mistaken for JSX and never hidden. The check uses the char-scan
  tokenizer, not the export reader's lexer, deliberately: that lexer's regex
  detector misreads the / of a </span> closing tag as a regex and would blank
  a real reference between two closing tags. Instead, regex-literal handling lives
  in the char-scan tokenizer with a JSX-safe guard — a / in expression position
  (including right after return/typeof/etc.) opens a regex and is blanked,
  EXCEPT when it immediately follows < (a </Tag> close or < /re/ comparison).
  So const r = /Foo/ stops false-firing while a / Foo / b, x > Foo, a
  variable named myreturn / Foo, and a real {Foo} between closing tags are
  never touched.