bug: silent fallback to full — a compressed ctx_read that returns the whole file says nothing

[yvgude]

Reported by a user running 3.9.20 (official build, path jail active, doctor green).

Symptom

Outside the project root, ctx_read(mode="signatures") returned the full file
with no indication that no compression had happened. From the caller's side a
lossy summary and a verbatim full read are indistinguishable — you pay full-file
tokens believing you asked for a summary. Same session: ctx_delta reported
"no cached version for diff" immediately after a successful mode=full read.

"A silent fallback to full is the bad failure mode — you pay full tokens
believing you asked for a summary."

Root cause

cap_to_raw (tools/ctx_read/core_logic.rs) enforces the #361 anti-inflation
invariant: if the framed/compressed output is larger than the raw file, the raw
file is returned instead. It returned it bare — no mode label, no note. The
same is true of raw_fallback in render.rs (the monotonic guard).

For a verbatim request (full, raw, lines:N-M) that is correct and silent
is right. For a request that asked for a summary (signatures, map,
aggressive, entropy, task, reference) silence is a lie by omission.

Fix (implemented)

A size-gated honesty banner:

[lean-ctx] no compression applied (mode=signatures): output was not smaller
than the file — returning full content (N tok)

• Emitted only when the requested mode counts_as_compressed().
• Emitted only above NO_COMPRESSION_BANNER_MIN_TOKENS = 400, so the banner can
  never itself push a small read above the raw size — Independent benchmark of lean-ctx on a pi/forge agentic workload — request for review before June 17 publication #361 keeps its guarantee
  ("a read never costs more than the raw file") exactly where it matters.
• Applied at both fallback sites (cap_to_raw and render::raw_fallback) so the
  CLI/in-process path and the MCP path behave identically.

Tests: cap_to_raw_labels_silent_fallback_for_compressed_request,
cap_to_raw_stays_silent_for_verbatim_full_request,
cap_to_raw_notice_never_breaks_the_361_cap.

Still open

The exact trigger for the out-of-root case was not reproduced end to end; a
/tmp vs /private/tmp path-normalization mismatch in the cache key is the
leading hypothesis and would also explain the ctx_delta "no cached version"
right after a successful read. Tracked separately if it recurs.

[yvgude]

Fixed — on branch fix/gh-issue-batch

You put it exactly right: "a silent fallback to full is the bad failure mode — you pay full tokens believing you asked for a summary." The output was indistinguishable from a summary that happened to need every line, and the failure surfaced only on the bill.

Root cause. Two places return the untouched file:

• render::raw_fallback — the monotonic guard: the compressed render came out no smaller than the source.
• core_logic::cap_to_raw — the Independent benchmark of lean-ctx on a pi/forge agentic workload — request for review before June 17 publication #361 anti-inflation cap: framing a whole-file view cost more tokens than the file itself.

Both returned bare content with no signal.

Fix. A single one-line banner, emitted by both paths:

[lean-ctx] no compression applied (mode=signatures): output was not smaller than the file — returning full content (1 240 tok)

Two deliberate boundaries:

• Only for compressed requests. Verbatim views (full, anchored, raw) stay byte-identical, so callers that demand exact bytes (compress-protected paths, chore(deps): bump fast-uri and @vscode/vsce in /vscode-extension #1150) never get a prefix.
• Only above NO_COMPRESSION_BANNER_MIN_TOKENS (400). Below that, the framing itself was the inflation — exactly what the Independent benchmark of lean-ctx on a pi/forge agentic workload — request for review before June 17 publication #361 cap exists to strip — and a banner would push the read back above the raw file the cap just protected. Independent benchmark of lean-ctx on a pi/forge agentic workload — request for review before June 17 publication #361's guarantee ("a read never costs more than the raw file") is preserved.

Regression tests

• render_tests::no_compression_banner_only_above_threshold — silent below the threshold, explicit above it, names the requested mode
• tests_inflation::cap_to_raw_labels_silent_fallback_for_compressed_request

The second half of your report — ctx_delta saying "no cached version for diff" right after a successful mode=full read — is the MCP-handler dispatch gap, fixed in #1584.

Not yet released: on fix/gh-issue-batch, not in 3.9.20.