15 — Verification pass and the extended matrix

Round-2 of the source program (2026-07-03): every load-bearing claim in this hub's RTK/headroom/lean-ctx pages was verified against fresh clones with file:line evidence, codedb was folded into the comparison matrix, and the "best part from each" selection was re-derived from the corrected picture. Caveman-family results are on 14 — The caveman ecosystem: flagship verified, cavemem and caveman-code deep-dived. Clones: rtk develop@b52e02b (no commits since), headroom c75ebde (no commits since), lean-ctx a80350da == v3.8.18 (no commits since), codedb c12e1e2 == v0.2.5826 (history squashed to one commit; no changes since 2026-06-23).

TL;DR — the corrections that matter

• The "one independent 47.5% headroom measurement" is dead. The figure appears nowhere outside headroom's own repo; its closest source is headroom's own "codebase exploration 47%" self-benchmark (41,254/78,502). The hub's long-standing "best-externally-evidenced tool" framing loses its only external datapoint. The 4.8%-median/50k+-sessions telemetry is real but historical and unrepeatable: the beacon that collected it was removed entirely (telemetry is now local-only, opt-in) — no newer aggregate can exist.
• lean-ctx counts, settled: 77 shell-pattern modules (74 routed), 46 hook-rewritten command prefixes — the hub's "56" (stale 3.4.x docs) and "81 patterns" (that's the MCP-tool count) were both wrong. Read modes: 13, not 10. Binary: 75.9 MiB measured from the v3.8.18 release artifact — the old 64.7 MB is stale low and upstream's "~17 MB" claim is unfounded. Lean proofs: 84 theorems counted (the paper's own total says 53), 0 sorry confirmed. Contracts: ~30, not 24. Tool-count marketing drifts 71/80/81 across surfaces; actual registry 82; pinned minimal profile = 5 tools, unpinned default advertises 11.
• RTK counts, settled: 77 CLI commands / 62 command modules / 63 TOML filters / 86 rewrite rules (hub said 42/64/85). The full 12-strategy taxonomy is now captured verbatim (below). Telemetry is consent-gated opt-in (GDPR flow), not unconditional phone-home — but the payload is 26 fields and low_savings_commands can leak up to 3 command tokens (can include a user argument), contradicting its own TELEMETRY.md. Tee is not failure-only: Failures(default)/Always/Never modes plus unconditional force-tee pagination hints. #582 and #886 are closed; #1155 is still open; a second permission-escalation class (#1213, compound-command segments) was found and fixed. RTK Cloud is website-only — not evidenced in-repo.
• headroom nuances: kompress-v2 is gated, not replaced (still the default fallback strategy; TextCrusher is opt-in default-off; the Rust live-zone dispatcher still no-ops PlainText). The net_mutation_gain cache-economics formula was corrected upstream (warm-case penalty over S+ΔT — the old form was systematically pro-mutation). The oat01 auth-mode bug is fixed. learn is LLM-distillation with a dry-run-default/--apply gate writing to CLAUDE.local.md (not shared CLAUDE.md). Memory: only 2 of 6 tables are agent-scoped, no schema migrations. New --lossless proxy mode ships format-native reversible compaction with a runtime round-trip self-check — the cleanest new mechanism in the sweep.
• codedb folded into the matrix (below): the only tool of the five with a persistent call/dep graph; also the fattest default output in its own shootout (2.99× the cheapest backend per call), 1.7 GB steady RSS at 6.3k files, and an install surface that is the hub's new trust-boundary counterexample — undocumented global PreToolUse hook blocking 10 shell commands in every project, self-granted mcp__codedb__* permissions, competitor hook-fronting, and telemetry that uploads to codedb.codegraff.com on session close (the prior "local NDJSON" description was too kind). Version correction: current is 0.2.5826, not 5827.

RTK: the 12 strategies, verbatim (finally captured)

From docs/contributing/ARCHITECTURE.md:199-308 — the taxonomy the hub cited for two rounds without enumerating:

New RTK mechanisms worth recording (all verified, all adoptable): the anti-confusion truncation contract (one [N more lines] end marker only — interleaved omission markers were parsed as code by agents and caused retry loops; kept+N==total is unit-tested); unless anti-swallow guards on TOML short-circuit rules; per-filter panic isolation (catch_unwind → raw passthrough — fail-open in both crash and quality dimensions with never_worse); content-sniffing auto-detection for piped input (1 KiB fingerprints, banner-anchored); extract_json_object (backward-walk brace-balancer for JSON islands in noisy stdout); transparent_prefixes (wrapper commands stripped before rule-match, re-prepended after); trust-before-load for repo-local filters (SHA-256 pin, read-once TOCTOU-safe, fail-closed, CI-env-gated override); hook integrity sidecars (5-state verification, version-header drift detection); global truncation-cap constants with a reduced() guard that can never empty a list; the ledger's self-measuring quality loop (parse-failure table + top-passthrough = missing-filter roadmap); cc-economics' honest dual valuation (blended CPT labeled UNDERESTIMATES, active CPT OVERESTIMATES).

headroom: verification detail

Confirmed exactly: live-zone five-part algorithm (frozen floor/ceiling/512 B thresholds/token-validation gate/RawValue surgery + manifest), CCR schema and marker grammar (Python accepts 12–24 hex), TTL 30 min now env-tunable, 1.25×/0.1× multipliers, bug-classes #1006/#1077/#1307 all fixed in Unreleased, org move with HF model still on the founder's personal account, Serena installed unpinned from git main (supply-chain anti-pattern). Newly extracted: TOIN = Tool Output Intelligence Network — observation-only local pattern store ((auth_mode, model_family, structure_hash) keys, hashed field names, no network I/O) with an offline toin_publish → recommendations.toml flow the Rust proxy loads at startup; toin_read_only=true for subscription traffic quarantines the learning pool from UX-bound accounts. Output shaper: byte-stable verbosity levels appended after the last cache_control breakpoint (deliberate cache constraint), effort downgrade only on structurally-classified mechanical continuations, conversation-stable holdout (sha256 cohorting) with a stratified normal-approximation CI estimator — but no auto-disable gate and no committed results. The relevance layer is linear BM25+embedding fusion with adaptive alpha driven by content signals (UUIDs→0.85 lexical, hostnames→0.6), not RRF. SharedContext remains a RAM dict (confirmed unchanged). The rtk data plane is a 5-second-timeout shell-out to rtk gain --format json whose result is collapsed via max(cli_filtering, rtk) — unattributable double-counting dodge, flagged do-not-copy.

lean-ctx: verification detail

Beyond the TL;DR corrections: benchmark numbers regenerated upstream (map 98.1%@78%, signatures 96.7%@96% — the hub's 97.8/77 and 96.5/95.9 are one refresh stale; conclusions unchanged); the CFT bandit is fully specified — 3 Beta(2,1) arms carrying (entropy, jaccard, budget) triples, success ⇔ savings-ratio > 0.3 AND task completed, ×0.98 session-end decay, and deterministic argmax by default with stochastic exploration behind an env flag, explicitly to keep provider prompt caches stable (#498) — the learning-must-not-destabilize-prefixes doctrine in code; bounce constants (5-tick window, ≥0.30 rate over ≥3 reads forces full, 10-tick post-edit exclusion so compression is never blamed for edit-forced re-reads); auto-mode default table (markdown/docs never compressed; learned results may never override to lossier than the static default); anchored read mode (per-line hash anchors making compressed-read→patch safe); Kolmogorov gzip-ratio gate before aggressive modes; footprint self-ablation (#959: each injected element A/B-ablated for pass-rate delta → prune_recommended, Ed25519-signed artifacts); doctor overhead --gate mechanics (runtime sum vs [context] budget_tokens; minimal-arm prefix budget pinned at 1,835 tokens; net_of_injection() can go negative); the eval testbench (same-budget arms, seeded deterministic bootstrap, record/replay, LLM-judge that hard-errors on unparsable verdicts); and one genuinely novel unrecorded subsystem — the scent field (stigmergic multi-agent coordination: 5 scent kinds with per-kind half-lives in one flocked JSON file, zero tokens; cites a 3.4× coordination-token claim). Security nuance: the A2A HMAC transport signs but the receiving endpoint never verifies — advisory only.

codedb: folded into the matrix

Best case: structural navigation of one large warm multi-session codebase — callers/callpath/transitive-deps in one ms-latency round-trip against a live graph none of the other four have, plus lifecycle (watcher, warmup, unchanged:HASH, hash-verified reader.md orientation, agent locks + ConvergenceGovernor anti-loop nudges). Hardest failure: plain text-search economics, per its own data — raw SQLite FTS5 used 23% fewer tokens end-to-end; wrong task + default config = schema rent + RAM + fattest-in-class outputs for something a sqlite3 one-liner beats. New adoptables: ranking-env fingerprint in cache keys; ConvergenceGovernor (8-slot call-signature ring, 3 repeats → nudge); warmup replay through the exact handler path; reader.md hardening set (traversal rejection, entry caps, LOC ceiling, golden hash-parity test, length-gated prepend); the failed-experiments ledger as a practice; two-stage mtime→hash change detection; per-response token cost as a first-class benchmark axis.

The extended matrix (seven tools, corrected)

Best-part-from-each, re-derived (the final selection)

The round-2 findings sharpen — and in two cases change — what the jackin❯ Context Engine spec takes:

The layered verdict itself is unchanged — output register + write-time input pipeline + structural retrieval + honest measurement, composed in one runtime-owned engine — but the evidence hierarchy shifted: headroom loses its only external validation, lean-ctx gains the most spec-grade mechanism detail, RTK gains the most portable safety contracts, and the caveman family contributes constants and counterexamples rather than mechanisms.

Ledger