drydock-cli 2.9.122__tar.gz → 2.10.1__tar.gz

{drydock_cli-2.9.122 → drydock_cli-2.10.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.9.122
+Version: 2.10.1
 Summary: Local-first CLI coding agent — tested with Gemma 4 26B via llama.cpp
 Project-URL: Homepage, https://github.com/fbobe321/drydock
 Project-URL: Repository, https://github.com/fbobe321/drydock

{drydock_cli-2.9.122 → drydock_cli-2.10.1}/drydock/cli/textual_ui/app.py

@@ -2055,6 +2055,21 @@ class DrydockApp(App):  # noqa: PLR0904
             await self._mount_and_scroll(UserCommandMessage(msg))
             return
 
+        if sub == "setup":
+            # Self-bootstrap — invokes the same logic as
+            # `python -m drydock.graphrag setup`. No-op safe to re-run.
+            import subprocess
+            import sys as _sys
+            res = subprocess.run(
+                [_sys.executable, "-m", "drydock.graphrag", "setup"],
+                capture_output=True, text=True, timeout=300,
+            )
+            output = (res.stdout or "") + (res.stderr or "")
+            await self._mount_and_scroll(UserCommandMessage(
+                f"### GraphRAG setup\n\n```\n{output.strip()}\n```"
+            ))
+            return
+
         if sub == "ingest":
             target = Path(rest or ".").expanduser().resolve()
             if not target.exists():

{drydock_cli-2.9.122 → drydock_cli-2.10.1}/drydock/core/prompts/gemma4_slim.md

@@ -32,6 +32,8 @@ WHEN TO USE retrieve (project's GraphRAG index):
 - "What does the project doc say about X?" — `retrieve(query="<topic>")`
 - Prefer `retrieve(query="ClassName")` over grepping the same file 5+ times for a symbol
 
+If the user asks to "set up GraphRAG" / "configure RAG" / "is the cookbook indexed?", run `bash` with `python3 -m drydock.graphrag setup` — it auto-detects bundled + local cookbooks and ingests them. No-op safe to re-run. To confirm afterward, run `python3 -m drydock.graphrag stats`.
+
 WHEN TO USE WEB TOOLS (sparingly):
 - Stuck on an error you've tried to fix 2+ times: `web_search` for the exact error message
 - Need an API example you don't remember: `web_search`

drydock_cli-2.10.1/drydock/data/cookbook/tbench_domains/biopython_dna.md

@@ -0,0 +1,116 @@
+# Bioinformatics / DNA assembly — reference for tbench tasks
+
+Used by: `dna-insert`, `dna-assembly`, `protein-assembly`. The tasks
+involve manipulating FASTA sequences, designing primers, simulating
+restriction-enzyme cuts, or Golden Gate / Gibson assembly. The
+standard tool is **Biopython** (`from Bio import SeqIO`).
+
+## Read / write FASTA
+
+```python
+from Bio import SeqIO
+from Bio.Seq import Seq
+from Bio.SeqRecord import SeqRecord
+
+# Read
+records = list(SeqIO.parse("sequences.fasta", "fasta"))
+for r in records:
+    print(r.id, len(r.seq))
+
+# Write — no blank lines, no extra whitespace
+out = [SeqRecord(Seq("ATGC..."), id="primer1_fwd", description=""), ...]
+SeqIO.write(out, "primers.fasta", "fasta")
+```
+
+Common gotcha: `SeqIO.write` adds a trailing newline. If the verifier
+checks "no blank lines", trim with `.replace("\n\n","\n")` after.
+
+## DNA basics
+
+```python
+from Bio.Seq import Seq
+s = Seq("ATGCTAG")
+s.complement()         # 'TACGATC'
+s.reverse_complement() # 'CTAGCAT'
+s.translate()          # protein, stop at TAA/TAG/TGA
+```
+
+## Melting temperature (Tm) — primer3-compatible
+
+The verifier often demands `primer3`'s `oligotm` output, NOT Biopython's.
+For matching ground truth:
+
+```python
+import subprocess
+def tm(seq):
+    out = subprocess.check_output(
+        ["oligotm", "-tp", "1", "-sc", "1",
+         "-mv", "50", "-dv", "2", "-n", "0.8", "-d", "500", seq],
+        text=True
+    )
+    return float(out.strip())
+```
+
+Biopython's `Bio.SeqUtils.MeltingTemp.Tm_NN` is close but NOT
+bit-identical to `oligotm`.
+
+## Restriction enzyme cut sites
+
+| Enzyme | Recognition | Cut produces |
+|---|---|---|
+| BsaI / BsaI-HF v2 | `GGTCTCN` (or rev `NGAGACC`) | 4-nt 5' overhang, cuts 1/5 downstream of GGTCTC |
+| BsmBI | `CGTCTCN` | similar Type IIS |
+| EcoRI | `GAATTC` | 5' overhang `AATT` |
+
+**Golden Gate assembly** uses Type IIS enzymes like BsaI. To make
+primers ready for one-pot Golden Gate:
+
+1. Add BsaI cut site to each primer's 5' end.
+2. Pick a 4-nt overhang (the "fusion site") that is unique across the
+   pool.
+3. Forward primer: `[stuffer 4-6nt] GGTCTC N [4-nt overhang] [template-annealing 18-30 nt]`
+4. Reverse primer: `[stuffer 4-6nt] GGTCTC N [4-nt overhang (complementary to next fragment's start)] [template-annealing rev_complement]`
+
+## Validation checklist for designed primers
+
+The verifier usually checks:
+- Each primer has BsaI site (`GGTCTC`) somewhere
+- Tm of the annealing portion in [58, 72] °C
+- Fwd/rev pair Tm within 5 °C
+- Annealing portion length in [15, 45] nt
+- Each fragment's overhangs match the NEXT fragment's start
+- No blank lines in the FASTA output
+- Header format `>TEMPLATENAME_DIR` where DIR is `fwd` or `rev`
+
+## Skeleton
+
+```python
+from Bio import SeqIO
+from Bio.Seq import Seq
+import subprocess
+
+BSAI = "GGTCTC"
+def design_primer(template_seq, overhang, length_anneal=22):
+    fwd_anneal = str(template_seq[:length_anneal])
+    rev_anneal = str(Seq(template_seq[-length_anneal:]).reverse_complement())
+    fwd = "AAAA" + BSAI + "A" + overhang + fwd_anneal
+    rev = "AAAA" + BSAI + "A" + overhang_rev_for_next_fragment + rev_anneal
+    return fwd, rev
+
+def tm(seq):
+    return float(subprocess.check_output(
+        ["oligotm","-tp","1","-sc","1","-mv","50","-dv","2",
+         "-n","0.8","-d","500", seq], text=True
+    ).strip())
+```
+
+## Common task pitfalls
+
+- **Always write the FASTA to the path the task names** — not a Python
+  script that produces it.
+- **Tm validation uses `oligotm`, not Biopython** — if you compute Tm
+  internally for primer design then output a primer with Tm 65°C, the
+  verifier will recompute and might see 67°C. Cushion the bounds.
+- **Overhangs must be UNIQUE in the pool** — if 2 primers have
+  overhang `AATG`, Golden Gate produces a mixed product. Use a hashing
+  function to assign each primer a unique 4-mer from a small palette.

drydock_cli-2.10.1/drydock/data/cookbook/tbench_domains/feal_cryptanalysis.md

@@ -0,0 +1,124 @@
+# FEAL cipher + cryptanalysis basics
+
+Used by: `feal-linear-cryptanalysis`, `feal-differential-cryptanalysis`.
+FEAL is a 64-bit block cipher with 32-bit half-blocks and a Feistel
+structure. The 4-round version (FEAL-4) is the standard
+cryptanalysis-teaching target.
+
+## FEAL-4 round function
+
+```python
+def f(half32, subkey32):
+    """FEAL round function: 32 bits in, 32 bits out, with a 32-bit subkey."""
+    a, b, c, d = (half32 >> 24) & 0xFF, (half32 >> 16) & 0xFF, (half32 >> 8) & 0xFF, half32 & 0xFF
+    ka, kb, kc, kd = (subkey32 >> 24) & 0xFF, (subkey32 >> 16) & 0xFF, (subkey32 >> 8) & 0xFF, subkey32 & 0xFF
+    b ^= a ^ ka
+    c ^= d ^ kd
+    b = G1(b, c ^ kb)          # G1, G2 = nonlinear 8-bit s-box variants
+    c = G0(c, b ^ kc)
+    a = G0(a, b)
+    d = G1(d, c)
+    return (a << 24) | (b << 16) | (c << 8) | d
+
+def G0(x, y):
+    return rol8((x + y) & 0xFF, 2)
+def G1(x, y):
+    return rol8((x + y + 1) & 0xFF, 2)
+def rol8(v, n):
+    return ((v << n) | (v >> (8 - n))) & 0xFF
+```
+
+## Encryption (FEAL-4)
+
+```python
+def feal4_encrypt(plaintext64, K1, K2, K3, K4, K5, K6):
+    L = (plaintext64 >> 32) & 0xFFFFFFFF
+    R = plaintext64 & 0xFFFFFFFF
+    L ^= K5
+    R ^= K6
+    R ^= L
+    for K in (K1, K2, K3, K4):
+        L, R = R, L ^ f(R, K)
+    L, R = R, L
+    R ^= L
+    return (L << 32) | R
+```
+
+## Differential cryptanalysis (high-level)
+
+Given enough (plaintext, plaintext') pairs with a specific difference
+`Δp`, observe the output difference `Δc`. For a "good" Δp the
+distribution of intermediate-state differences is non-uniform — the
+top key candidates have heavier counts.
+
+Standard FEAL-4 differential attack:
+- Choose Δp such that the input difference cancels after the first 3
+  rounds with high probability (≈ 1).
+- Generate ~10,000 chosen-plaintext pairs.
+- For each pair, guess the last-round subkey, partial-decrypt, check
+  if the resulting difference matches expectation.
+- The correct key has the highest hit count.
+
+## Linear cryptanalysis (high-level)
+
+Find a linear approximation `(plaintext_bits XOR ciphertext_bits) =
+key_bit` that holds with probability != 1/2. Matsui's piling-up lemma
+chains per-round approximations.
+
+For FEAL-4, a working 3-round approximation exists with bias ~2^-5.
+Need ~2^10 known plaintexts to recover one key bit reliably.
+
+## Tools that may already be installed
+
+- `pycryptodome` (`from Crypto.Cipher import ...`) — does NOT include
+  FEAL by default; you'll need to implement it
+- Numpy for vectorized XOR / counting across many pairs
+
+## Common task shape
+
+The task gives a known FEAL-4 encryption oracle and asks you to recover
+the full key (6 × 32 bits = 192 key bits) using only chosen-plaintext
+or known-plaintext queries.
+
+A working approach:
+1. Implement FEAL-4 encrypt + decrypt exactly per the round function above.
+2. Run a per-byte differential search on the last round: try all 256
+   values of each subkey byte against ~5000 pairs with Δp = 0x80800000.
+3. Once last-round subkey is known, peel off and recurse for round 3, 2, 1.
+4. The first two key registers (K5, K6) are recovered from the
+   pre-whitening XOR at the start — fixed by 2 plaintexts with known
+   relationship.
+
+## Skeleton
+
+```python
+import struct, random
+
+def encrypt_oracle(pt: int) -> int:
+    # Provided by the task — call as a black box
+    ...
+
+def differential_attack(num_pairs=20000):
+    delta = 0x80800000_00000000  # canonical FEAL-4 diff
+    counts = [[0]*256 for _ in range(4)]  # one per byte of K4
+    for _ in range(num_pairs):
+        p1 = random.getrandbits(64)
+        p2 = p1 ^ delta
+        c1 = encrypt_oracle(p1)
+        c2 = encrypt_oracle(p2)
+        for k_byte_idx in range(4):
+            for k_guess in range(256):
+                # partial decrypt last round using k_guess for this byte
+                # check if internal-state difference matches Δ_expected
+                ...
+    # top-counted guess per byte = correct key byte
+    K4 = bytes(max(range(256), key=lambda k: counts[i][k]) for i in range(4))
+    return int.from_bytes(K4, "big")
+```
+
+## Pitfalls
+
+- The byte order matters: FEAL-4 spec is big-endian. If your bytes
+  come out wrong, swap endianness.
+- The "+1" in `G1` is easy to forget — it's what makes G1 != G0.
+- `rol8` is a *byte* rotation, not 32-bit — verify with tests.

drydock_cli-2.10.1/drydock/data/cookbook/tbench_domains/mips_isa.md

@@ -0,0 +1,162 @@
+# MIPS ISA + ELF parsing — reference for tbench tasks
+
+Used by: `make-mips-interpreter`, `make-doom-for-mips`, `extract-elf`.
+Tasks usually give a compiled MIPS ELF binary and ask for an emulator
+that runs it (often in JavaScript so it runs in node).
+
+## MIPS instruction encoding (32-bit, big-endian or little-endian per ELF)
+
+3 formats, all 32 bits:
+
+```
+R-type (register):  | opcode(6) | rs(5) | rt(5) | rd(5) | shamt(5) | funct(6) |
+I-type (immediate): | opcode(6) | rs(5) | rt(5) | imm(16)                     |
+J-type (jump):      | opcode(6) | target(26)                                  |
+```
+
+Decode from a 32-bit word `w`:
+
+```python
+opcode = (w >> 26) & 0x3F
+rs     = (w >> 21) & 0x1F
+rt     = (w >> 16) & 0x1F
+rd     = (w >> 11) & 0x1F
+shamt  = (w >> 6)  & 0x1F
+funct  = w & 0x3F
+imm    = w & 0xFFFF        # I-type, sign-extend for arithmetic
+target = w & 0x03FFFFFF    # J-type, shifted left 2 + bottom 4 bits of PC+4
+```
+
+## Top opcodes to implement
+
+| Mnemonic | opcode | funct | What |
+|---|---:|---:|---|
+| `add`/`addu` (R-type) | 0 | 0x20/0x21 | rd = rs + rt |
+| `sub`/`subu` | 0 | 0x22/0x23 | rd = rs - rt |
+| `and/or/xor/nor` | 0 | 0x24/25/26/27 | bitwise |
+| `sll/srl/sra` | 0 | 0x00/02/03 | rd = rt shifted by shamt |
+| `sllv/srlv/srav` | 0 | 0x04/06/07 | rd = rt shifted by rs |
+| `jr` | 0 | 0x08 | PC = rs |
+| `jalr` | 0 | 0x09 | rd = PC+8; PC = rs |
+| `mult/multu/div/divu` | 0 | 0x18/19/1A/1B | sets HI:LO regs |
+| `mfhi/mflo` | 0 | 0x10/12 | rd = HI or LO |
+| `slt/sltu` | 0 | 0x2A/2B | rd = (rs < rt) ? 1 : 0 |
+| `addi/addiu` (I-type) | 0x08/0x09 | — | rt = rs + sign_ext(imm) |
+| `andi/ori/xori` | 0x0C/0D/0E | — | rt = rs op zero_ext(imm) |
+| `lui` | 0x0F | — | rt = imm << 16 |
+| `lw/sw` | 0x23/0x2B | — | word load/store |
+| `lb/lbu/sb` | 0x20/24/28 | — | byte load/store |
+| `lh/lhu/sh` | 0x21/25/29 | — | half-word |
+| `beq/bne` | 0x04/0x05 | — | branch if rs ==/!= rt |
+| `blez/bgtz` | 0x06/0x07 | — | branch on sign |
+| `j` (J-type) | 0x02 | — | PC = (PC & 0xF0000000) \| (target << 2) |
+| `jal` | 0x03 | — | $ra = PC+8; jump |
+| `syscall` | 0 | 0x0C | OS hook (file I/O, exit) |
+
+## Register file
+
+32 general-purpose regs, plus HI/LO for mul/div results, plus PC.
+By convention:
+
+```
+$0  ($zero)   always 0 — writes silently ignored
+$1  ($at)     assembler temp
+$2-$3 ($v0-1) return values
+$4-$7 ($a0-3) function args
+$8-$15 ($t0-7) temp regs
+$16-$23 ($s0-7) saved regs
+$24-$25 ($t8-9) more temp
+$26-$27 ($k0-1) kernel
+$28 ($gp)     global pointer
+$29 ($sp)     stack pointer
+$30 ($fp)     frame pointer
+$31 ($ra)     return address
+```
+
+## SPIM/Linux syscalls
+
+`syscall` looks at `$v0` for the syscall number; args in `$a0-$a3`:
+
+| $v0 | What | $a0 | $a1 |
+|---:|---|---|---|
+| 1 | print_int(int) | int | — |
+| 4 | print_string(addr) | C-string addr | — |
+| 5 | read_int → $v0 | — | — |
+| 8 | read_string(buf, len) | buf addr | len |
+| 10 | exit | — | — |
+| 11 | print_char | char | — |
+| 4001 | Linux write | fd | buf addr | (with $a2=count, returns in $v0)
+| 4003 | Linux read | fd | buf addr | (with $a2=count)
+| 4004 | Linux open | path addr | flags |
+| 4006 | Linux close | fd | — |
+| 4001-ish | Linux exit | code | — |
+
+## ELF parsing (MIPS32 little-endian common)
+
+`fitz`/`pyelftools` not always available; manual reader:
+
+```python
+import struct
+
+def parse_elf(path):
+    with open(path, "rb") as f:
+        d = f.read()
+    assert d[:4] == b"\x7fELF"
+    # 32-bit, little-endian
+    e_entry      = struct.unpack_from("<I", d, 0x18)[0]
+    e_phoff      = struct.unpack_from("<I", d, 0x1C)[0]
+    e_phentsize  = struct.unpack_from("<H", d, 0x2A)[0]
+    e_phnum      = struct.unpack_from("<H", d, 0x2C)[0]
+    segments = []
+    for i in range(e_phnum):
+        off = e_phoff + i * e_phentsize
+        p_type, p_offset, p_vaddr, p_paddr, p_filesz, p_memsz, p_flags, p_align = \
+            struct.unpack_from("<8I", d, off)
+        if p_type == 1:  # PT_LOAD
+            segments.append({
+                "vaddr": p_vaddr,
+                "filesz": p_filesz,
+                "memsz": p_memsz,
+                "data": d[p_offset : p_offset + p_filesz],
+            })
+    return e_entry, segments
+```
+
+## Memory model for the emulator
+
+Allocate ~16MB virtual memory as a flat `Uint8Array` (JS) or `bytearray`
+(Python). Load each segment at `vaddr`. Stack at `0x7FFFFFF0` growing
+down. Reads/writes via `dataview.getInt32(...)`.
+
+## Minimum interpreter skeleton (Python — port to JS line-for-line)
+
+```python
+class CPU:
+    def __init__(self, mem, entry):
+        self.r = [0] * 32
+        self.hi = 0; self.lo = 0
+        self.pc = entry
+        self.mem = mem  # bytearray of size N
+
+    def load_word(self, addr):
+        return struct.unpack_from("<I", self.mem, addr)[0]
+
+    def step(self):
+        w = self.load_word(self.pc)
+        opcode = (w >> 26) & 0x3F
+        next_pc = self.pc + 4
+        if opcode == 0:
+            funct = w & 0x3F
+            rs, rt, rd, shamt = (w>>21)&0x1F, (w>>16)&0x1F, (w>>11)&0x1F, (w>>6)&0x1F
+            if funct == 0x20 or funct == 0x21:  # add/addu
+                self.r[rd] = (self.r[rs] + self.r[rt]) & 0xFFFFFFFF
+            # ... rest of R-types
+        elif opcode == 0x09:  # addiu
+            rs, rt = (w>>21)&0x1F, (w>>16)&0x1F
+            imm = w & 0xFFFF
+            if imm & 0x8000: imm -= 0x10000  # sign-extend
+            self.r[rt] = (self.r[rs] + imm) & 0xFFFFFFFF
+        # ... rest
+        self.r[0] = 0  # $zero always 0
+        self.pc = next_pc
+```

drydock_cli-2.10.1/drydock/data/cookbook/tbench_domains/path_tracing.md

@@ -0,0 +1,110 @@
+# Path tracing / ray tracing — reference for tbench tasks
+
+Used by: `path-tracing`, `path-tracing-reverse`. The task usually asks
+you to render a scene to a PPM/PNG given camera + sphere geometry, OR
+to recover scene parameters from a rendered image.
+
+## Core math
+
+### Ray-sphere intersection
+
+A ray `R(t) = O + t * D` (origin O, direction D, parameter t >= 0)
+hits a sphere `|P - C|^2 = r^2` (center C, radius r) when:
+
+```
+let L = O - C
+a = dot(D, D)        # = 1 if D is normalized
+b = 2 * dot(L, D)
+c = dot(L, L) - r*r
+disc = b*b - 4*a*c
+if disc < 0: no hit
+sqrt_disc = sqrt(disc)
+t1 = (-b - sqrt_disc) / (2*a)
+t2 = (-b + sqrt_disc) / (2*a)
+# nearest positive root is the hit
+```
+
+### Lambertian (matte) bounce direction
+
+Cosine-weighted hemisphere sample at hit point with normal N:
+
+```python
+import random, math
+
+def lambertian_bounce(N):
+    # Sample uniform direction on unit sphere
+    while True:
+        x, y, z = random.uniform(-1,1), random.uniform(-1,1), random.uniform(-1,1)
+        if x*x + y*y + z*z < 1.0:
+            break
+    # Project into hemisphere aligned with N
+    v = (x, y, z)
+    if dot(v, N) < 0:
+        v = (-v[0], -v[1], -v[2])
+    return normalize(v)
+```
+
+### Tone mapping + gamma
+
+Linear color → sRGB (8-bit) by gamma 2.2:
+
+```python
+def linear_to_srgb_8bit(x):
+    x = max(0.0, min(1.0, x))
+    return int(255 * (x ** (1/2.2)))
+```
+
+### PPM output (P3 ASCII or P6 binary)
+
+```python
+# P3 ASCII — easiest to debug
+with open("out.ppm","w") as f:
+    f.write(f"P3\n{W} {H}\n255\n")
+    for r,g,b in pixels:
+        f.write(f"{r} {g} {b}\n")
+
+# P6 binary — smaller, byte-exact
+import struct
+with open("out.ppm","wb") as f:
+    f.write(f"P6\n{W} {H}\n255\n".encode())
+    for r,g,b in pixels:
+        f.write(bytes([r,g,b]))
+```
+
+## Common task shapes
+
+**Forward render**: given camera origin, view direction, fov, sphere
+list — produce W×H image. Loop over pixels, for each cast a ray, hit
+the nearest sphere, return its color (possibly with N bounces of
+Lambertian sampling for diffuse global illumination).
+
+**Reverse**: given a rendered PPM + partial scene, recover the missing
+parameter. Usually involves binary-searching the missing scalar
+(sphere radius, light intensity, etc.) by re-rendering and comparing
+L2 norm against the reference.
+
+## Numerical stability gotchas
+
+- **Self-intersection**: when reflecting at hit point P with normal N,
+  offset the new ray's origin by `P + 1e-3 * N` to avoid hitting the
+  same surface at t=0.
+- **Negative disc check**: always test `disc < 0` BEFORE `sqrt(disc)`.
+- **Normalize D once per ray**: don't recompute `dot(D,D)` every hit.
+
+## Minimum skeleton
+
+```python
+def render(W, H, spheres, eye, fov_deg):
+    import math
+    f = 1.0 / math.tan(math.radians(fov_deg) / 2)
+    pixels = []
+    for j in range(H):
+        for i in range(W):
+            # NDC -> camera space
+            x = (2 * (i + 0.5) / W - 1) * (W / H) / f
+            y = (1 - 2 * (j + 0.5) / H) / f
+            d = normalize((x, y, -1))
+            color = trace(eye, d, spheres, depth=0)
+            pixels.append(tuple(linear_to_srgb_8bit(c) for c in color))
+    return pixels
+```

drydock_cli-2.10.1/drydock/data/cookbook/tbench_domains/vim_macros.md

@@ -0,0 +1,120 @@
+# Vim macros — keystroke-efficient text transformation
+
+Used by: `large-scale-text-editing`. Task asks for `apply_macros.vim`
+that transforms an input CSV / text file to match a target using only
+Vim normal-mode keystrokes recorded into registers a, b, c, with a
+keystroke budget (usually <200 total). Run headlessly:
+
+```
+vim -Nu NONE -n -Es input.csv -S apply_macros.vim
+```
+
+## Vim script structure for this task
+
+```vim
+" Define 3 distinct macros — keystrokes stored as register strings
+call setreg('a', "<keystrokes>")
+call setreg('b', "<keystrokes>")
+call setreg('c', "<keystrokes>")
+
+" Apply each to every line (or globally)
+:%normal! @a
+:%normal! @b
+:%normal! @c
+
+" Save and quit
+:wq
+```
+
+## Keystroke encoding for setreg()
+
+In a string, special keys must be escaped:
+
+| Want | Encode as |
+|---|---|
+| `<Esc>` | `\<Esc>` (or `^[` literal — best avoided) |
+| `<CR>` (Enter) | `\<CR>` |
+| `<Tab>` | `\<Tab>` |
+| `<C-x>` | `\<C-x>` |
+| `<BS>` (backspace) | `\<BS>` |
+
+Wrap the string with double quotes (not single) so Vim parses
+backslashes.
+
+## Ex commands inside a macro
+
+`:%normal!` applies normal-mode keystrokes to every line.  Macros
+can also include `:` commands by recording them with `:` + Enter inside
+the keystroke string:
+
+```vim
+call setreg('a', ":s/foo/bar/g\<CR>")
+```
+
+That's "type colon, type the substitute, hit Enter".
+
+## Keystroke-efficient patterns
+
+| Goal | Compact keystrokes |
+|---|---|
+| Delete to end of line | `D` (= `d$`) |
+| Replace whole line | `cc` |
+| Replace single char | `r<char>` |
+| Append at end of line | `A` |
+| Substitute on every line | `:%s/PAT/REP/g<CR>` |
+| Insert column at start | `:%s/^/PREFIX/<CR>` |
+| Reorder CSV fields | `:%s/\v(\w+),(\w+),(\w+)/\3,\1,\2/<CR>` |
+| Skip first line (header) | `:2,$normal! @a` instead of `:%normal! @a` |
+| Repeat last action | `.` |
+
+## Counting keystrokes
+
+The task counts characters in the recorded register, including
+escape sequences. `\<Esc>` is 5 chars, NOT 1. To stay under 200 total:
+
+- Prefer `:%s/X/Y/g<CR>` over per-line normal-mode loops
+- Use `^` to anchor at start of line, `$` at end — avoids `0` and `$`
+  motions
+- Use `&` to repeat the last substitution on a different selection
+
+## Worked example — uppercase first column of CSV
+
+Input:
+```
+foo,1,a
+bar,2,b
+```
+Output:
+```
+FOO,1,a
+BAR,2,b
+```
+
+One-macro solution (47 chars in register):
+
+```vim
+call setreg('a', ":%s/^\\(\\w\\+\\)/\\U\\1/<CR>")
+:normal! @a
+:wq
+```
+
+`\U\1` = uppercase the first capture group.
+
+## Three-macro template
+
+Split the transformation into three independent passes — that's what
+the task usually expects. E.g.:
+
+- `a`: normalize whitespace
+- `b`: re-quote strings
+- `c`: re-order columns
+
+Each pass is one substitute or one `:%normal!` invocation.
+
+## Common pitfalls
+
+- Single quotes in `setreg` don't process `\<Esc>` etc. — USE DOUBLE QUOTES.
+- `:wq` exits with status 0 only if file was modified or read-only
+  wasn't set. Use `:x` if the verifier insists on no-op safety.
+- `vim -n` disables swap file (required for headless reliability).
+- `vim -Nu NONE` skips vimrc loading — your macros run in vanilla Vim.