fix(read): adopt upstream #3377 — --max-lines as exact contiguous head count

[kylehgc]

Closes #50.

What this adopts

Adoption of upstream rtk-ai/rtk#3377 by Mohammed Elkarsh (@mohammedelkarsh): replaces smart_truncate with a plain contiguous head_truncate in read.rs, making --max-lines N (the hook's head -N mapping) return the literal first N content lines plus the [K more lines] marker.

Why rtk-ai#3377 instead of the ticketed rtk-ai#2964 + rtk-ai#3166

Issue #50 named upstream rtk-ai#2964 (exact budget) + rtk-ai#3166 (contiguous window), which conflict with each other by construction and disagree on whether the marker counts against the budget (rtk-ai#2964: N content lines; rtk-ai#3166: N−1). During Phase 2 the upstream sweep surfaced #3377 (opened 2026-08-03), which supersedes both in one coherent change:

• Exact budget (fix(read): fill the --max-lines budget exactly instead of stopping at half rtk-ai/rtk#2964's bug): exactly N content lines, marker extra — head -N parity.
• Contiguous prefix (fix(read): head -N truncation must be a literal contiguous prefix rtk-ai/rtk#3166's bug): plain slice, no importance heuristic pulling a bare } or export from past the window.
• Plus --max-lines 0 → empty output (head -0 parity), and it deletes the now-unused smart_truncate outright instead of patching it.

Adopting rtk-ai#3377 is one clean cherry-pick with no hand-authored conflict resolution; adopting the ticketed pair would have required rewriting rtk-ai#3166's diff during the pick. Verified: smart_truncate's only caller was read.rs:182; the FUNC_SIGNATURE/IMPORT_PATTERN regexes remain in use by AggressiveFilter (no dead code).

Commits

Commit	Author	Kind
653fb56 fix(read): honor --max-lines N as exact head count	Mohammed Elkarsh	Cherry-pick, byte-identical to upstream 91bec76
34bc096 test(read): pin head window to a literal contiguous prefix (rtk-ai#3046 shape)	kylehgc	Fork amendment

Amendment rationale: rtk-ai#3377's tests use plain a..f lines, so a re-introduced importance heuristic would pass them. The amendment reproduces the original rtk-ai#3046 corruption shape directly (60 plain lines, bare } at line 61, 30-line window).

Repro (debug binary, Windows x64)

Before (develop @ d08ba98) — both defects in one view:

$ rtk read plain10.txt --max-lines 4        # 10 plain lines
line1
line2
[8 more lines]                              # asked for 4, got 2

$ rtk read brace.txt --max-lines 30         # 62 lines, bare } at line 61
alias thing_1='do stuff 1'
...
alias thing_15='do stuff 15'
}                                           # pulled from line 61, false adjacency
[46 more lines]                             # 16 content lines for a 30-line ask

After (this branch):

$ rtk read plain10.txt --max-lines 4
line1
line2
line3
line4
[6 more lines]

$ rtk read brace.txt --max-lines 30 | tail -2
alias thing_30='do stuff 30'
[32 more lines]                             # 30 contiguous lines, no phantom }

Quality gate

cargo fmt --all (no-op) · cargo clippy --all-targets 0 warnings · cargo test --all 2641 passed / 0 failed across 18 suites (x86_64-pc-windows-msvc host toolchain).

Removed tests: the five smart_truncate unit tests died with the function (upstream's own diff); their invariant (kept + reported == total) is re-asserted by the adopted test_head_truncate_plain_text_exact_count and the fork amendment.

Deliberately not fixed

• Upstream fix(read): bound memory for --tail-lines instead of loading the whole file rtk-ai/rtk#3132 (bound memory for --tail-lines) — separate concern, same file; left for its own ticket. Note it will conflict textually with this change's apply_line_window signature cleanup if adopted later.
• Upstream fix(read): fill the --max-lines budget exactly instead of stopping at half rtk-ai/rtk#2964 / fix(read): head -N truncation must be a literal contiguous prefix rtk-ai/rtk#3166 stay open upstream, untouched — no comments left there. Issue Adopt upstream #2964 + #3166: read - head-rewrite fidelity (exact budget, contiguous window) #50's note to re-check old upstream fix(read): preserve exact head semantics rtk-ai/rtk#1018 for residual value: fix(read): honor --max-lines N as exact head count rtk-ai/rtk#3377 delivers exact head semantics, so fix(read): preserve exact head semantics rtk-ai/rtk#1018 is superseded on this fork.

🤖 Generated with Claude Code

[kylehgc]

Review round (pre-merge, /code-review two-axis)

Standards axis: no hard violations. Judgement calls noted: (1) no ≥60% savings assertion — deliberate, this is an explicit-detail flag path where CONTRIBUTING.md's correctness-over-savings principle governs; (2) synthetic 3000-line fixture in the adopted exact-count test — acceptable per cli-testing.md inline-fixture guidance; (3) minor two-instance collect/slice/join duplication between the tail path and head_truncate — extraction would cost more than it saves; (4) deleting the unused _lang param and heuristic path flagged as a positive (Speculative Generality removed).

Spec axis: fully implemented. Independently verified the cherry-pick 653fb56 is byte-identical to upstream 91bec76 and the amendment 34bc096 is tests-only. One nuance recorded: on truncation the output joins with \n (CRLF files lose \r, no trailing newline before the marker) — identical to the pre-existing tail path, so consistent module semantics, not a regression.

No changes required; merging as reviewed.