format

• Breaking for code written against 3.0.0: canFormat, canLookup, and canRepresent now receive the extension's T; the engine checks the runtime type before invoking user code, and each predicate accepts every T by default. Remove overrides that only returned value is T; rewrite an additional filter from Object? to T.
• Avoided converting ordinary int values to BigInt while validating Unicode scalars for {:c} and %c. In a local dart2js A/B on arm64, cached {:c} fell from 133 ns to 56 ns and %c from 108 ns to 29 ns; Format 1.6 took 143–144 ns for the brace case. Real BigInt values retain the same validation path and showed no reproducible regression.

Upgrading from the published 1.6.0 also includes the changes of the unpublished 2.0.0 below.

• Added a brace-formatting engine using the Python mini-language, with positional, named, item, and attribute lookup; conversions; nested fields; and typed errors.
• Added sprintf and vsprintf with a C-style subset for text, integer, decimal floating-point, and hexadecimal floating-point conversions.
• Added Dart SDK and Python/C++-compatible decimal double profiles. Dart SDK conversion is the default; compatible mode preserves nearest-even rounding, exponent layout, extended precision, and inf/nan spellings.
• Added configurable NaN/Infinity or nan/inf spelling in Dart SDK mode, and applied the selected profile recursively to !r and !a conversions.
• Represent empty Dart Map and Set values as {}.
• Added immutable Format instances with configurable custom formatters, lookups, representations, number locales, and Unicode text units.
• Bounded the template cache by memory as well as by entry count: templateCacheMemoryLimit (8 MiB per mini-language by default) and templateCacheMemory join templateCacheCapacity and templateCacheSize. Whichever bound binds first evicts, an entry priced above the whole budget is formatted but not cached, and lowering either bound discards entries immediately. The figure is an estimate — a Dart program cannot measure the memory it holds — priced from the template text, the slices its literals keep, the code units prepared for them, and a constant per parse node, with the constants fitted to measured retention. Bytes rather than characters because the same amount of text costs from 1 to 154 bytes per character depending on how densely it is fielded, which is what a character count cannot see. An entry is repriced when an engine of the other Unicode text unit reaches it: the parse is shared, but the compiled program and the specifications memoized under it are held per unit, and that second copy measures between a quarter and three fifths of the first price.
• A template with no fields — or a printf template with no conversion — is now formatted without copying it: it compiles to a single op that hands the text back by reference. Measured on the VM at 2.6 times faster for a sixteen character template and 4600 times for one of a hundred thousand, where the copies were the whole cost, and a cached entry for such a template retains about a third of what it did.
• A template that is a single padded field — {:>12.2f}, {:^10s}, %10s — assembles its result as one string instead of writing the fill around the body into the buffer and building the string back out of it. Where the field is the whole output that round trip buys nothing. Measured on the VM: {:>12.2f} 179 ns becomes 140, {:012.2f} 181 becomes 145, {:^12.2f} 189 becomes 152, {:>10s} 83 becomes 50, %10s 86 becomes 56 and %12.2f 179 becomes 144. The web backends accumulate into a string to begin with and so move less — under dart2js {:>12.2f} 107 ns becomes 104, under dart2wasm 150 becomes 144. Templates of more than one field are unchanged, and so is a field whose sign, percent suffix or grouping separator has to be written around or through the body rather than beside it.
• Integers beyond 2^53 no longer go through BigInt to reach base 2, 8 or 16 on the web. Every radix this package supports is a power of two, and a binary double converts into one exactly, so the platform's own conversion already spells the value. Measured under dart2js at 2^53: {:x} 320 ns becomes 120, {:o} 4070 becomes 150, {:b} 11550 becomes 380. Decimal is unchanged — it still needs fixed-point conversion, and BigInt past 1e21.
• A specification with a nested field — {value:{width}.{precision}f} — is parsed once per resolved text rather than once per call. The resolution itself still happens every call, because it is part of the values and not of the template, and a resolution that changes is parsed again. Measured under dart2js: {0:>{1}.2f} 366 ns becomes 257 and {0:{1}d} 191 becomes 175; under dart2wasm 497 becomes 341, on the VM 558 becomes 490.
• Grouped integer conversions on the web no longer count the digits by dividing: the platform conversion that produces them already knows how many it wrote. Measured under dart2js on a sixteen-digit value: {:,d} 272 ns becomes 200 and {:020,d} 262 becomes 182; under dart2wasm 215 becomes 199 and 214 becomes 198. The VM writes its digits into a buffer and is unchanged.
• The template cache stops consulting itself when it is being thrashed, and starts again after a while in case the workload has changed. What triggers it is misses in a row that each had to evict something — a cache filling up is all misses too, and that is not the same thing — and what it holds is kept rather than discarded. This is the regime the README already documented as the one where caching never pays at any repetition rate. Measured on a first call under dart2js: a literal template 600 ns becomes 250, one of ten fields 4230 becomes 2240; on the VM 469 becomes 311 and 2337 becomes 2043; under dart2wasm 379 becomes 237 and 2424 becomes 1900. A workload that does repeat is unaffected — its hits are what tells the two apart — and templateCacheSize still reports what is held.
• The g, e and bare-precision floating presentations take their digits from the platform's own exponential conversion instead of decomposing the double and rounding in BigInt. The exact path stays for what the platform cannot answer: past twenty-one significant digits, and on the values that round to an exact half, where the SDK and ECMAScript round away from zero while this package rounds to even. Output is unchanged everywhere — checked against exact rounding on 313 200 comparisons per runtime, where the only disagreements were those ties. Measured under dart2js: {:g} 1437 ns becomes 132, {:e} 1113 becomes 323, %g 1473 becomes 132, {:.6} 1462 becomes 147. On the VM {:g} 448 becomes 266, under dart2wasm 457 becomes 205. A value that does land on a tie pays for the attempt before falling back and costs about a third more than before, 1.7 times under dart2wasm.
• Fixed-point conversion keeps the platform's own spelling past the range where the scaled value is still an exact double, where it used to fall back to BigInt. Rounding ties are decided from the bits of the value there, which is exact at that range and needs no scaled product. Measured under dart2js: {:.6f} of 12345678901234.568 goes from 1849 ns to 86, and {:.2f} of 1.23e19 from 2070 to 87; on the VM 694 to 158 and under dart2wasm 606 to 94. Values inside the old range are untouched, and their cheaper arithmetic test is still what decides them.
• formatWith snapshots its named arguments as a plain copy instead of an unmodifiable view of one. The snapshot still guarantees that a toString reached during the call cannot change what that call reads. Measured on a template of three named fields at 4% faster on the VM, 16% under dart2js and 18% under dart2wasm, where every named lookup had been going through the view.
• Lowered the characters constraint to ^1.3.0. Flutter pins that package from its SDK, and stable releases pinning 1.3.0 could not resolve this package at all (issue #8). The two versions carry different Unicode grapheme tables — 15.0.0 and 16.0.0 — so TextUnit.graphemeClusters follows whichever version resolves; nothing else in the package depends on the difference, and CI now runs the whole test suite on the floor to keep that true.
• A fraction separator may now be written without a precision — '{:.,f}' — which CPython's grammar allows and this package rejected. It groups the fraction at the presentation's own default precision, so '{:.,f}' of 1234.5678 is 1234.567,800 and '{:_.,f}' of 1234567.5678 is 1_234_567.567,800, both exactly as CPython writes them. A lone . is still an error, and an integer conversion still has no fraction to separate.
• Breaking, from the published 1.6.0: n no longer follows intl. 1.6.0 depended on intl and read Intl.defaultLocale, so '{:n}' localized itself and '{:,n}' was accepted. 3.0.0 drops that dependency: n follows the configured NumberLocale, which is the C locale unless one is set, and '{:,n}' is rejected — a locale decides its own grouping, so asking for a separator and for the locale's separator at once has no answer. An application localized on 1.6.0 keeps compiling only if it sets a locale, and gets the companion package below for the intl data it used to inherit.
• Added the companion format_intl package for opt-in locale symbols, grouping rules, and localized digits without coupling format to intl.
• Extended the configured NumberLocale to the printf integer conversions, which previously took their sign and digits from ASCII while %f and %e already read the locale. %d, %i, %u, %o, %x and %X now use the locale's signs and digits, padding zeros included; the 0x marker of %#x stays ASCII, and the alternate zero of %#o localizes as the digit it is. Only digits are localized, so the hexadecimal letters of %x and %X pass through unchanged and a locale with non-ASCII digits produces both scripts in one number — %x of 0xabc123 under Eastern Arabic digits is abc١٢٣. The same rule governs %a, where the digits and exponent localize while 0x and p do not. Nothing changes under the default C locale.
• Cache parsed templates, and expose templateCacheCapacity, templateCacheSize, and clearTemplateCache() so an application whose working set is larger than the default can widen it, one whose templates never repeat can switch it off, and either can tell the two cases apart.
• Exported TextUnitOperations, so a configured TextUnit can measure and truncate text the same way the engine does.
• Defined consistent JavaScript number semantics and optimized decimal integer formatting, including large values and the minimum VM integer; decimal digits beyond 2^53 print exactly on the web instead of the JavaScript shortest or exponential forms.
• Rejected widths and printf options above 100000 with typed errors instead of attempting arbitrarily large allocations.
• Added cross-runtime compatibility fixtures and a reproducible JIT, AOT, and JavaScript benchmark harness, measured against the frozen Format 2 gate baseline and the published sprintf 7.0.0 and format 1.6.0 packages.
• Added an ANSI-colored example benchmark that displays the result and timing of both decimal double profiles side by side.
• Replaced formatNamed with formatWith; direct format and sprintf calls now accept up to ten values, while their collection-based counterparts are formatWith and vsprintf.
• Formatting exceptions render their type, payload, and full template context in toString(), and a value whose own toString() throws is reported safely. FormatExceptionContext.fragment is an excerpt rather than a copy: it is capped at 80 characters, ending in … and never cut inside a surrogate pair, while template is kept whole. Use offset with template when the exact span matters.
• Documented the whole public API surface with dartdoc, including the extension contracts an implementer otherwise had to read the engine for: what a throwing canFormat, format, lookup, or represent turns into, that built-in types take priority over an extension, and that a Map uses the string-key shortcut instead of a registered AttributeLookup.

• Upgrade depencencies.
• Min sdk: 3.3.0.