Chameleon

A second same-day real-usage QA round, aimed at the surfaces v2.50.0 shipped
after its own QA passes closed: the class/module name search, the Python
module_attribute call edges, and the refresh envelope cap — plus a
whole-plugin day-in-the-life pass on Flask, Next.js, and Rails repos. All three
v2.50.0 additions survived hostile testing (the relative-import join held
against parent-package decoys, the envelope cap is accurate and lossless, all
five sampled honesty fixes behave as claimed, and search line numbers were
exact on every spot check across the three languages). The gaps this release
closes came from the edges of those features and from dogfooding chameleon on
its own large source files.

Added

• Absolute-import phantom-symbol check for Python. from flaskbb.utils.helpers import totally_fake — a real in-repo module, a
  hallucinated name, the single most common LLM import mistake — previously
  passed the post-edit verify as clean, because only relative imports were
  scanned and repos whose own idiom is absolute imports (most Flask and Django
  apps) effectively had no symbol check at all. The scan now also resolves
  absolute first-party specifiers against the repo's Python source roots and
  flags a bound name absent from the resolved module's closed export set. The
  module itself is never flagged: an unresolvable specifier may be stdlib or a
  dependency, so it stays silent; open export sets (__getattr__, star
  re-exports) are respected; and a from pkg import name where name exists
  on disk as the package's submodule file or subpackage directory is never
  flagged even when the index misses it (PEP 420 namespace subpackages are
  unenumerable at dump time, and an index built by an older engine may
  predate submodule listing — reality on disk beats the index).
• Ruby block-form nested classes are searchable by their qualified name.
  module Qa51Mod; class Qa51Nested was findable only by leaf name while the
  compact form (class Qa51Mod::Qa51Nested) carried its full path — the same
  concept, findability depending on definition spelling. The Ruby dump now
  records an additive qualified constant path (lexical nesting join) and the
  symbol index prefers it, so qualified queries hit, leaf queries still hit on
  the substring tier, and two same-leaf nested classes (BudgetCategory::Group
  vs Category::Group) render disambiguated instead of as identical rows.
• Search results carry kind and honest truncation. search_codebase
  rows now say what they are (class vs the callable kind), so "find class
  Account" is distinguishable from same-named attribute readers, and a query
  matching more symbols than the result cap sets truncated plus a note
  instead of silently dropping the tail. describe_codebase god-symbol counts
  saturated silently at the per-callee stored-row cap (a symbol with 177 real
  callers displayed as 100, disagreeing with search_codebase on the same
  name); a capped count now rides out with capped: true so it reads as a
  floor.

Fixed

• module_attribute false edge on class-member names. mod.method()
  where method exists only INSIDE a class in the target module recorded a
  call edge to a dispatch that is an AttributeError at runtime — the gate
  checked the target's full callable table, class members included, unlike the
  Ruby grade's member-kind gate and the TS closed-export gate. The gate now
  requires a module-level definition. Zero legitimate edges lost on the real
  corpora (the Django fixture's 67 edges survive exactly).
• Phantom "removed-export" violations on files over 100KB. The post-edit
  scan reads a capped 100KB prefix; when that prefix happened to parse as
  valid Python, every export defined past the cap read as removed — editing
  chameleon's own conventions.py produced 5 bogus "removed-export"
  violations naming present symbols (and hook_helper.py produced 27). The
  removed-export check is now skipped on truncated content (absence from a
  truncated prefix is not evidence of removal); importer-count advisories
  keep applying to visible names.
• Transitive caller walk no longer burns its budget on duplicate call
  sites. The calls index stores one row per call site, and the shared
  blast-radius/judge walker expanded a full duplicate subtree per site — in
  chameleon's own index 18.8% of caller rows are duplicate sites and 34% of
  callees are affected, so truncated walks hid real unique callers behind
  repeats (and the turn-end judge live-reverified the same chain twice).
  Expansion now dedups per node by caller function: threshold_int's reach
  went from 41 to 48 distinct callers under the same caps.
• Editing an archetype's canonical witness no longer injects that file's
  own content back as the thing to imitate. On a Flask repo whose view
  witness is a 1,000-line module, editing it produced a 43KB Tier-2 block
  whose "canonical witness" was the edited file itself, led by "mirror the
  canonical witness below closely." A self-witness edit now gets a one-line
  note (keep the exemplar's conventions stable) instead, and any witness
  excerpt is capped at a line boundary (TIER2_WITNESS_MAX_CHARS, default
  16000) with an honest truncation marker, so one pathological canonical
  cannot dominate every edit's block.
• Class contracts are measured over the dominant cohort, not a rich
  minority. A decorator carried by 4 of an archetype's 76 classes could
  clear the file-count anchor gate (Flask view modules pack dozens of classes
  per file) and out-rank the real contract on richness — every view edit
  then read "decorated with @attr.s; define get, post, redirect" as archetype
  fact when 72 of 76 classes disagree. Candidate anchors must now cover a
  comparable share of the largest candidate cohort before competing on
  richness; the Flask fixture's contract corrected to extends MethodView; define post, get over all 76 classes.
• Rails model-migration advisory skips POROs and stops repeating. A plain
  class under app/models (no ActiveRecord descent) was told to add a
  migration it cannot need; the rule now requires
  ApplicationRecord/ActiveRecord::Base descent in the added file's
  content. The same still-unresolved advisory also re-rendered verbatim on
  every consecutive Stop; it now surfaces once per session per (file, rule),
  the same discipline the idiom self-review and finding ledger follow.
• TypeScript class definition lines and class-expression names. A
  decorated class reported its first decorator's line (every NestJS class was
  offset from the class keyword, inconsistently with Python); the name
  identifier's line is recorded now. A named class expression (const C = class Inner {}) was indexed under the body-scoped Inner — a name repo
  code cannot use — and an anonymous-but-bound expression not at all; both
  now index under the variable binding.
• Smaller honesty items. get_canonical_excerpt's no_witness reason
  claimed "below confidence threshold, or all candidates contained secrets"
  when the usual cause is the canonical-pool exclusion of test/legacy paths —
  the reason now names the real causes. doctor's recent_hook_errors warn
  is labeled installation-wide so a fresh repo's user does not read another
  repo's three-day-old fixture errors as their problem. The Tier-2
  nearby-signatures section logged nothing when its assembly failed (a vanished
  section was indistinguishable from "no siblings"); the swallowed exception
  now lands in the hook error log. The v2.50.0 CHANGELOG's "false-positive-
  free" claim for module_attribute was corrected to name the documented
  binding-shadow imprecision and the class-member gap fixed above.