readability_db.md

Readability carry-forward

Three generations of readability/ have existed. All three are gone from the working tree; all three are recoverable from git.

• First generation, 84 files and roughly 6,700 lines, distilled 2026-08-09 and deleted in 8660e7c1. Recover any file with git show 8660e7c1^:readability/~30_Patterns--Observer.r2.md.
• Second generation, 47 chapter files written 2026-08-11 (7f165a08, "readability on the whole deep-review") over the fresh deep-review round. Every finding was applied to Chapters/ at the time, though the files were not renamed with the ~ completion prefix until 2026-08-13. Distilled into this file 2026-08-13 and deleted. Recover any file with git show bb81711a:readability/14_Techniques--Decorators.md, with no ~ on the path: the rename and the deletion landed in the same change, so the committed second-generation files carry the bare chapter name even though every one of them had been applied.
• Third generation, 92 files written 2026-10-16, one per chapter and one per Solutions README. Distilled 2026-10-16 into "The 2026-10-16 generation" at the end of this file, which gives the recovery command.

What a fresh readability pass should do with this file. Read it first. This skill hunts AI-writing tells and watch-list vocabulary, which makes it the pass most likely to re-flag something a previous run already examined and kept. Everything below survived that examination. Re-raising any of it is churn.

---

Standing rejection

08_Foundations--Static_Types: keep the two sentences after "the same idea checked at different moments". The review proposed cutting "Dynamic typing trusts the object once the code is running, while structural typing proves the shape beforehand" as §70 restatement, since the paragraph states the timing contrast three times. Declined. Both sentences stay.

---

The recurring keeps

The second generation's declined lists run to roughly 200 entries, and almost all of them are instances of the dozen judgments below. A mechanical sweep re-finds every one. Read this section as the answer to "why does the book keep doing X", not as a list of exceptions.

A watch-list absolute stays when it states a guarantee. never, ever, only, already, even are the words this pass flags most, and the book keeps them wherever the word carries a universal or temporal claim rather than emphasis. The test: does weakening it to "does not" trade a guarantee for a report? If so, keep it. "Assignment never copies" is a claim about every case; "hands out references, never copies" is the contrast that explains the leak; "a lazily imported name nobody touches never loads" is the silent-side-effect point; "Circle and Square never mention Drawable" is a fact about the source. The same reasoning keeps "only" wherever it marks a real restriction ("Only the checker sees it, and only at edit time" is two parallel restrictions, not filler) and "already" wherever it is temporal or marks work the reader does not have to do.

"want" addressed to the reader is the carve-out. The don't-use entry targets anthropomorphized objects ("the function wants a string"). "A bug you want flagged", "the broad catch you want instead", "if you really want many handles sharing one set of state" all address the reader's intent and stay. This precedent is cited by a dozen chapters.

"is what" survives when its complement cannot attach without it. The global rule states its own carve-out and the book leans on it: "The yielded value is what as binds", "f3() is what forgetting looks like", "The rest is what SimpleNamespace adds", "The counts, not a stopwatch, are what this listing measures". Apply the deletion test; where deleting breaks the sentence rather than tightening it, stop.

"worth" stays where the frame weighs a real cost. §53 flags the endorsement family, and its carve-out covers most of the book's uses: "a separate factory class is worth writing when …", "the hand-rolled Messenger is worth writing only to show …", "Blob is the case worth watching:". Each weighs an effort against a stated condition, with the payoff delivered in the same sentence. (2026-10-04, Sweep Decisions: Bruce split the Blob sentence at its colon; the "worth" stays. Chapter 11's "worth the small effort" became "justifies the effort", his wording.)

§34 real/actual/genuinely/truly stays where the contrast is named. "the caller's copy of the list still holds your actual Bobs", "It uses genuinely separate functions per type", "When a program truly needs an object", "the real class that does the work". In each the sentence itself names what the word contrasts with, which is the rule's own exception.

"plain" stays against a named alternative. Against defaultdict, against a generator function, against a descriptor, against a bound method, against an async function. The book uses it as a contrast marker, not a filler qualifier.

"itself" stays when reflexive and load-bearing. "The list automatically resizes itself", "A sub-pattern is itself a pattern" (without it the sentence is a tautology), "a container holding an unhashable object is itself unhashable" (the word marks propagation).

A rhetorical question stays when the question is the content. §43 targets stalling transitions. "How does with know what to run?", "What should the type annotation be?", "What happens if gather() encounters a failure?" each name the genuine problem the section then solves, and each is its chapter's only one.

Rule-of-three stays when the prose then uses the items. §10 flags the shape; the book's triples are usually inventories the following sentences pick apart one at a time ("Tally.total is the first of these. For the third…"). Padding to reach three is the defect, not three real items.

Predicate-position hyphens stay on terms of art. §26 would drop them, but thread-safe, black-box, I/O-bound, type-blind, positional-only, and keyword-only are hyphenated in standard usage regardless of position, and several are terms the chapter defines and its headings use.

A single Tier 1A or Tier 2 vocabulary hit is not a cluster. The tiers fire on density. "daunting", "myriad", "compelling", "valuable", "Notably", "Certainly" each survived as the sole occurrence in their chapter, inside first-person or first-edition passages.

A short emphatic fragment is a beat, not a staccato run. "It has a sharp edge.", "No class, no ceremony.", "seen is how.", "I hardly do." Each sits among long sentences and does deliberate work. (2026-10-04, Sweep Decisions: Bruce cut two fragments that stated a verdict on output already on the page, chapter 18's "Same answer, from 242,785 calls against 26." and chapter 47's "Four implementations, one Ability, one running program."; the keep covers a beat, not a verdict.)

An expletive "there is" belongs to /activate, not here. "Unlike C, there is no fall-through", "For event-heavy programs there are mature libraries". The register pass owns that category; flagging it in a readability pass duplicates work and invites conflicting rewrites.

"hook" is the literal term, not a metaphor. Chapter 17 uses it throughout (including a hook_order.py listing and a "Which Hook for Which Job" heading), chapter 25 introduces it in italics as GoF's term, and chapter 26 names __getattr__() the fallback hook, which exercise 4 then depends on. Renaming it anywhere breaks terminology with the chapter that owns it.

---

Conventions confirmed clean

Recorded so a sweep does not read them as defects:

• Colon reveals in 40_Functional--Foundations ("A match is code: … The table is data: …") are §69 by shape, twice. Both colons introduce a definition rather than staging a surprise, and the parallel structure is the point of the contrast. The same call covers 06's "The underscore changes one mechanical thing:" and 02's "not a matter of taste: it is the structure".
• Contrastive italics introducing a thing in its new role: "is a module", "is a closure", "is a hook", "is a handler", the "what"/"how" pair. Book-wide convention across chapters 14, 25, 28, 29, 40, and 44, not italics-for-emphasis.
• "which is why" as the causal connector. Chapter 06 alone has seven. Each ties a listing's observed output to the mechanism behind it, which is the book's standard move. Only genuinely adjacent pairs were varied.
• Conditions, not imperative-plus-consequence. "If you delete either total = 0, the second assertion fails" was deliberately written as a condition rather than the banned "Delete either … and …" form. Several such rewrites landed during the 2026-08-09 annealing pass across chapters 31, 36, 40, 41, 42 and 47; they are the corrected form, not candidates for reversal.
• Imperatives in exercises are exempt. Exercise text instructs the reader, which the global rules carve out. Do not convert exercise imperatives to conditions.
• Parallel prose decision tables. 35_Patterns--Flyweight's four-way If X, do Y block and 26_Patterns--Factory's bullet list under "Which Factory Should You Use?" are decision tables written as prose; the uniformity is what makes them scannable. (2026-10-04, Sweep Decisions: within that uniformity Bruce cut 35's trailing "which is the least machinery for the job" as awkward, and gave 27's "When a name maps to a class" bullet a comma, matching its siblings.) 24_Patterns--Template_Method's four-way "Structure fixes it… Discipline fixes the rest" anaphora is the same call.

---

Considered and declined, still binding

These are watch-list hits and pattern matches a mechanical sweep will find again. Each was judged and kept.

26_Patterns--Factory, "Factory might be the most common design pattern." An unsupported superlative by shape. Kept: it is hedged with "might", it opens the chapter, and it is first-edition voice. 11_Techniques--Testing's "One of the most valuable habits in modern programming is unit testing", 01_Introduction's "one of the most useful things this book can give you", and 21's "An important step forward in object-oriented design" all ride on this precedent.

27_Patterns--Function_Objects, "Python's best-known closure trap." A superlative by shape, but "best-known" orients the reader toward the trap they have heard of, and the late-binding loop variable is the one closure trap with a reputation.

27_Patterns--Function_Objects, "to say what one list of functions says directly." Not a finding at all: recorded as the sharpest sentence in the chapter and the reason the Command section works. Do not "tighten" it. (Superseded 2026-10-04, Sweep Decisions: Bruce, "sentence can be simpler"; the "directly" went and the sentence was simplified at his ruling.)

30_Patterns--Observer, "It is simply a callable." "Simply" is an empty adverb by the deletion test, but this is the deflating beat in the two-sentence pair carrying the chapter's central move, and the word supplies that spoken rhythm. 11_Techniques--Testing's "The 4 here is simply what Random(0) produces first" and 09's "annotations that merely describe one to come" are the same call.

30_Patterns--Observer, "amounts to nothing more than a list of callbacks." On the watch list, but the global rule's own carve-out names this exact construction: a comparative where the diminishing is the point. The rule's illustrative example is this sentence. 20's "costs nothing more than having the three methods" rides on it.

30_Patterns--Observer, "a lambda equals only itself." Both watched words are load-bearing: the claim is identity-equality, and "a lambda equals itself" without "only" is trivially true of everything. 22's "equals only its own kind" is the same shape. (2026-10-04, Sweep Decisions: 30's stays, with "bound to the same lambda" nearby at his note; 22's sentence he ruled "too long and complex" and had split, the restriction kept in words. He also had the "only" dropped from 42's "None says only 'no answer'" and 46's "only its signature" sentence rewritten as shorter sentences, so the watch-list- absolute keep does not cover those two.)

31_Patterns--State_Machines, "The conditions and actions are plain methods." "Plain" draws a real contrast: against the Condition and Transition class hierarchies the Java version needed, named a few paragraphs earlier.

33_Patterns--Visitor, the ## The Price of the Empty Base heading covering two paragraphs. The heading names only the first paragraph. Retitling (## What the Classic Shape Costs) and splitting were both considered and declined: the second paragraph ends on the line that motivates the whole singledispatch section, so it sits last regardless; a retitle makes the heading vaguer, and a new heading adds an anchor for two paragraphs. (Retitled 2026-10-13 to ## Two Departures from the Classic Shape, which names both paragraphs; the commerce rule R2 settled the "Price" question.)

33_Patterns--Visitor, "The output above shows results, not mechanism." Flagged as §70 metadiscourse. Declined by the block's own criterion: "mechanism" is established book vocabulary, thirty-plus uses across the chapters.

33_Patterns--Visitor, "a framework you do not own already calls that…" "Already" is on the avoid-if-possible list but earns its place: the call site exists whether you want it or not. This is the precedent a dozen other "already" keeps cite.

25_Patterns--Surrogate, "in order to implement the copy-on-write idiom." The one Tier 1B "in order" in the book, and the deep review deliberately restored it: the bare comma version read as two unrelated purposes, losing the causal link that counting references is how copy-on-write knows when to copy. Binding.

40_Functional--Foundations, "Choose match when the set of cases is fixed … and a table when the set is meant to grow from outside." §70 by shape, since the two sentences above draw the same distinction concretely. Kept: the abstraction is short, it is the sentence a reader returns for, and what precedes it are examples rather than a statement of the rule.

---

What was deliberately not carried

• Applied-directly lists, the bulk of both generations. Each entry describes an edit already in Chapters/. Verified by spot-check across chapters 03, 14, and 44 before the second generation was deleted.
• Chapter-local declined judgments that are instances of the recurring keeps above rather than separate rulings.
• Live blocks, which were applied when Bruce handed each file back.
• !Notes.md. Both directories had one, in both generations. deep_review/!Notes.md was an empty scratch file; readability/!Notes.md held only the skill's own how-the-workflow-works explainer, then nothing at all. Neither contained anything of Bruce's, so nothing personal was lost.
• The r1/r2 split (first generation only). Most chapters had two rounds; r2 findings supersede r1, and only r2's declined items were durable.

2026-10-04, Sweep Decisions page: cost tallies Bruce had cut. Chapter 20's "Satisfying three of them costs nothing more than having the three methods" and chapter 22's "The price of an ad-hoc attribute bag is that ..." lost their cost frames at his ruling (R2); 22 keeps the words "no type checker knows your attribute names" verbatim for the 22 -> 33 thread in deep_review_db.md. The twelve keeps he confirmed on that page (01's and 11's and 21's superlatives, 03's "assignment never copies", 06's underscore colon, 08's structural-typing restatement, 20's "truly", 31's "stored directly", 33's "results, not mechanism", 34's three-walker paragraph, 40's two-colon parallel and its "Choose match" sentence) stand as recorded.

---

The 2026-10-16 generation

The third generation reviewed the whole book on 2026-10-16, one Opus agent per file. It held 92 files: 48 chapter and appendix reviews named readability/<stem>.md, and 44 Solutions reviews named readability/<stem>.solutions.md. Together they list 650 direct fixes and 833 declined items, counted as list entries; a few entries bundle several hits of one kind. Eight files carried one block each. Bruce rejected none. All eight were applied the same day, which gave those files a leading ~. ~31_Patterns--State_Machines.md, a 2026-10-15 review of chapter 31 by the same pass (two direct fixes, five declines, no blocks), was deleted with them. The applied-directly lists stay out of this file, for the reason given above.

Recover any of these files with git show <hash>^:readability/~44_Effects--Effect_Management.md, where <hash> is the commit "Readability: distill the 2026-10-16 generation and clear readability/", the one that deleted them. git log --diff-filter=D -1 --format=%h -- readability/~44_Effects--Effect_Management.md prints that hash. The nine closed files keep the ~ in their recovery path; every other file has the bare stem.

The eight blocks, all applied

None of these is a standing rejection. Each sentence was decided on 2026-10-16, so a fresh pass treats the current wording as settled.

• 10_Foundations--Cleanup, the last sentence of "The Rule". "the leak it exists to catch" gave the registry a leak-detection purpose, where the chapter presents it as a live-instance counter. It now reads "so the registry leaves each object's lifetime as it was."
• 10 Solutions, exercise 4. "The registry now causes the leak it exists to detect" became "The registry now leaks every Counter it tracks," decided together with the chapter's sentence.
• 20_Patterns--Rethinking_Objects, line 140. The sentence now reads "Substitutability is the part of polymorphism that no tool can check." The proposed wording went in, over the alternative that made substitutability a separate guarantee.
• 21_Patterns--Design_Patterns, after the Design Principles list. The first-edition send-off "You can hold this handful of fundamental ideas in your head while analyzing a design" was cut, so the section ends on its last principle.
• 23_Patterns--Singleton, the mutate-versus-rebind passage. The middle statement of three, "Mutation makes the sharing work. Rebinding is the mistake that quietly ends it.", was cut. The figure paragraph, the code walk, and the refrain "Mutate through any name. Rebind only through the module." stay.
• 25_Patterns--Surrogate, the weakref cross-reference paragraph. Cut whole, since the paragraph above it closes the topic.
• 29 Solutions, exercise 3. The causal "so" before @staticmethod is gone, and the passage now reads "None of these functions uses self. @staticmethod makes Python call each one with the caller's arguments alone, even through an instance."
• 44_Effects--Effect_Management, line 75. The "because" clause now gives the reason in place of the definition: "because the operation that changes the environment sits in that function's body."

Keeps this round confirms

Most of the 833 declines are instances of the recurring keeps above. A sweep that re-finds them should stop at the same place. The 2026-10-16 instances of each:

• A watch-list absolute that states a guarantee. 08's "The Python runtime never acts on a type hint", 42 Solutions' "so it never propagates to the top-level handler", 34's "even with a case for every node" (the concession is the claim).
• "only" marking a restriction. The six shapes ruled fine on 2026-10-04 account for most only hits, among them 05 Solutions' "Only clear_by_method() empties the caller's list", 35's "Only identity proves sharing", and 41 Solutions' "accepts only an integer" (an open-ended alternative).
• "want" for the reader's intent or a human subject. 16 Solutions' "whether you want the list", 32 Solutions' "Which behavior you want", 37's "the safety officer wants".
• "is what" with a complement that needs it. 03 Solutions' "The rest is what Counter supplies", 46 Solutions' "That is what structural typing means". A cleft with a noun-phrase complement, repeated to answer an exercise's question, stays too, as in 29 Solutions' "What you lose is the log."
• "worth" weighing an effort against a condition. 07 Solutions' "is worth writing when a tool reads it", 27 Solutions' "The second assertion is the one worth writing", 47's "worth retrying", where retrying is the program's own vocabulary.
• real/actual/genuine with the contrast named. 19's "A real deadlock has no timeout" (against the demo's bounded run), 45's "The # type: ignore marks a real mismatch", 47's six uses of "real clock" against the handled clock.
• "plain" against a named alternative. 30 Solutions' "A plain except" against except*, 37's "an alias for a plain dict" against defaultdict, B's "a plain assignment" against the annotated one.
• A reflexive "itself". 07's "keeps each method from calling itself", 47's "It prints nothing itself" (against the supplied Narrator), 33's idiom "the end in itself".
• A rhetorical question that names the section's problem. 15's "What does with call?", 19's "How does gather() respond to a failure?", 21 Solutions' "What stops working? Nothing.", which repeats the exercise's own question.
• A rule of three that is an inventory. 13 Solutions' "an is test like this one, a membership test, or a call", 34's "SymPy, Polars, and SQLAlchemy" (the next sentence uses each), 36's "rewind, rollback, or restore".
• A single tier hit. 13's "valuable", 21's "One of the most compelling motivations", 01's "daunting" and "myriad".
• A short beat. 30's "It is not.", 32's "It works.", 43's "looks safe. It is not." 19's "Three unrelated kinds of worker produce one set of answers." was weighed against the 2026-10-04 cut of verdict fragments and kept, because it is a full sentence stating the section's point beside a listing with no #: markers.
• A superlative that orients the reader. The Factory precedent and 28's "best-known closure trap" cover 02's "one of the main reasons for Python's consistent readability", 04's "the classic control-flow bug", and 41's "The best-known case is Kirkman's schoolgirl problem".
• The book's own term, generalizing the "hook" keep. A watch-list word or figure stays when it names a chapter's term or a listing, as in 25 Solutions' near_miss.py (a listing name that holds a don't-use phrase), 39's "dissolve in Python" (chapter 21's heading "When a Pattern Dissolves"), and 43's "spectrum" and "rung", which the heading and the numbered list share.

New recurring keeps

These judgments recur across files and are new to this file.

• A negation that is the claim, or half of a named contrast. The positive-statement rule keeps these, and §9 fires on the "not X but Y" shape. 09's "The bug is not the class attribute. It is writing one where you meant a per-object default.", 33's "Nothing edits Flower.", 18's "The goal is not the fastest possible program." A silent failure takes a negative too, as in 16's "prints nothing and raises no exception".
• An additive "and" across a change of subject. The split rule fires on the subject change. The substitution test decides. Where no "but", "so", "while", or "whereas" fits, the "and" stays: two links of one chain (29's "Turning the key primes the pump, and priming starts the engine"), joint premises of a following "so" or "because" (07 Solutions' "The lookup starts at the class of the object, and self is a MoreDerived, so ..."), or two outcomes of one run (46's "Terminal prints Alice's greeting and capture stays empty"). Two splits were tried and reverted because they left a run of short sentences: 03's "in and len() work on all of them, and slicing works on the sequences" and 13's "Cases do not fall through ..., and match has no break to forget".
• Bold step labels are the Solutions convention. §15 and §16 fire on every Solutions README. The labels walk a solution's steps in prose (commit 85f6ce4d), as in 02 Solutions' "Tell values from literal text." and 05 Solutions' "Accumulate in the shared default.". A first sentence that echoes its label and then goes further stays, as in 17 Solutions' "Inherit the base's arithmetic. Celsius inherits float's arithmetic", which continues to the result type. A sentence whose whole content is its label was cut (see the last subsection). The same call covers the bold-label bullets of each chapter's Guidelines section (19) and A's numbered list, whose bold leads are full sentences.
• A repeat that a collapsed step needs. The hint, the shape, and the solution open one at a time, so a reader may see one step alone. A fact stated in the hint and again in the solution stays, as in 12 Solutions' 2000 as the one century year Year accepts, 35 Solutions' "record() forwards slots=False alone", and 47 Solutions' "Without the guard, each worker builds another pool."
• The trap-first opening. A Solutions section opens on the wrong version failing, and the labeled steps then walk the repair. The overlap is the file's shape, not a treadmill. This round's instances are 02 Solutions' solutions 1 and 5, 12 Solutions' solutions 2, 3, and 8, and 34 Solutions' RecursionError in exercise 2.
• A restatement that adds a step. §70 fires on the shape. The keep holds in three cases: when the second statement names the misreading (05's "The default looks like an expression each call evaluates, but Python evaluates it once, at the def."), applies the rule to a new subject (23's Collection requirement applied to Countdown), or re-establishes a frame after a digression (34's "Interpreter is Composite applied to language" after the pathlib digression).
• A refrain or a parallel repeated by design. 18's "The 3x threshold sits far below any multiple you should see." in both the NumPy and the Numba sections, 44's "An Effect Management System replaces ..." closing three separate arguments, 29's introduction previewing a section that other chapters link by its anchor, so the section restates the point for a reader who arrives there directly.
• An aphorism that states the mechanism. §32 fires on the shape. The keep holds when the sentence is concrete and true of the listing beside it, as in 22's "An immutable type needs immutable fields.", A's "A row that lists everything tells you nothing.", and 46 Solutions' "a value nobody uses is a value nobody checks". 44's "a space heater with extra steps" is a joke that carries the claim.
• A colon that passes the "namely" test. §69 fires on the shape. The colon rule keeps each of these: 12's "Nothing objects: not the annotation, not the type checker, not the running program" (the style guide's own example), 13's "It also has a name: the expression problem", A's "says how long that took: a decade".
• A count lead-in. 31's "Two consequences follow from the mixin", 04 Solutions' "Two things follow from that control.", 46's "Two details in greeter.py matter:". Each count is exact. The items follow at once.
• First-edition and first-person voice. Beyond the superlatives above, this round kept 17's "It is a fascinating tool and tempting to use", 25's "At the heart of a framework is the Template Method", and 21's "(I'll readily admit that I may be missing something here)". Epigraphs belong here (16, 28, 45, and 10, whose different-subject "and" stays as epigraph voice). The author's "we" stays wherever it appears (27, 30, 31, 44).
• The author's framing of common practice. §5 fires on an unsourced generalization. The book keeps these as the author's claims: 32's "Most programmers first write a third version.", 34's "Textbooks usually present the Interpreter pattern as ...", 44 Solutions' "Most codebases avoid the forwarding with a module-level logger."
• An adverb that carries content. §23's deletion test fails when the adverb states something the sentence needs, as in 22's "automatically creates" (Python builds the dict with no code from the class), 38's "works entirely in rooms" (no grid coordinates), and "silently" or "quietly" marking an error the reader would expect (12, 16).
• A commerce word for a measured quantity or the program's money. R2's carve-out covers 03 Solutions' "the whole loop costs O(n²)", 36's "cost about two million pointers", and 13's "a rough cost to send a message", money in the program's domain.
• "exactly" for a precise match. 19 Solutions' "Exactly half the increments survive" (200 of 400), 41's "exactly once", a combinatorial count, three times, 39's "exactly one receiver".
• Past tense for an event on the page. 38 Solutions' "the default kick was down to 0.00 by step 400", 45 Solutions' "The type checker told you more.", 46 Solutions' "greet_all() did the right thing all along."
• A sharper rewrite that would add a claim. The vaguer sentence stays when every sharper version states a fact the file lacks, as in 04's "the KeyboardInterrupt you press" (naming the key), 12's "Copying instead of mutating reduces errors.", and 24 Solutions' "where ruff expects every import".
• A paragraph's position answers the question its listing raises. Moving the paragraph was weighed and declined in three cases: 10's WeakSet aside, which answers "why a dictionary?" right after the listing, 32 Solutions' Origami(Paper) paragraph, and 19 Solutions' paragraph about the return in solution 10.

Conventions confirmed clean, 2026-10-16

• Review naming. A Solutions README review is readability/<chapter stem>.solutions.md, beside the chapter's readability/<chapter stem>.md. The readability skill names the chapter form alone. A completed file takes the ~ prefix on either name.
• Bold step labels and the six ruled-fine only shapes held across all 92 files.
• Hand-wrap width. Solutions READMEs are hand-wrapped near 70 columns, and an edited paragraph is rewrapped to that width. Hint paragraphs carry one sentence per line, in every Solutions file. Long or ragged lines left by an earlier edit, most often where a bold label went in, render the same, so a pass rewraps them when it edits the paragraph or when a line broke after one or two words.
• "The test file confirms that:" is the book-wide lead-in for test bullets (chapters 11, 23, 28, 29, 30, 34, 35, 36, and B).
• "Here," and a colon introducing a listing. B's four "Here," openers and 42's listing colons point at the listing that follows.
• A chapter roadmap in an opening paragraph (04, 40, 41, 45) orients the reader, so §28 and §70 leave it alone.
• Exercise wording. Exercise imperatives and questions stay, including 23's "What has first() cost you on an infinite source?", whose cost is the memory the listing reports.

Superseded, confirmed, and left open

• 26's "in order to implement the copy-on-write idiom" keep, above, has lapsed. The smart-reference item now reads "which makes copy-on-write possible", with the causal link intact.
• 08's standing rejection was confirmed again.
• Three declines were left for an editing pass rather than ruled on: 47 Solutions' "The same distribution is why the chapter uses a factory function rather than more supply() arguments" (the reason that follows is the nine-actor ceiling), 42's sentence pairing a deleted e with sum_type.py's kept results, and 23's closing list naming a peekable wrapper that exercise 8 builds.

What a fresh pass should still look for

The 650 fixes clustered on a few shapes. A fresh pass will find more of each. Roughly 170 split an "and" that joined clauses with different subjects where "but", "so", or "while" fit. Two keeps in 15 Solutions pass the "whereas" test, "Entering is outside-in ..., and exiting is inside-out ..." and "A truthy value suppresses the exception, and a falsy value lets it continue", so a fresh pass should split them too. About 100 cut a restatement, most often a Solutions sentence that repeated its own bold label ("Make the hidden input a parameter." followed by "Move the hidden input into the parameter list."). Smaller clusters trimmed empty intensifiers (a bare "real", "truly", or "really"), "worth" frames that rated a list in place of introducing it ("The cases worth testing are:"), and pointers with two candidates, where a "This" or an "it" could name either of two nouns in the sentence before. Checking each rewritten claim against its listing turned up a few inaccuracies along the way, so a pass that rewrites a sentence reads the listing first.