grattanInflators

• grattanInflators_has_no_data() checks for nonempty files without reading their contents; selecting an index still validates snapshot freshness.
• Inflator wrappers report invalid index classes through index validation instead of treating them as missing data.

Bug fixes (memory safety)

• Inflate(x = ) no longer writes past the end of x when x is shorter than from, and no longer requires x to be a double: the length is checked in both R and C, and an integer x is coerced (and so must be read back from the return value).
• The monthly, quarterly and annual kernels now bounds-check every index position before reading it. An out-of-range date gives NaN. This check is unconditional: check = 0 suppresses diagnostics, never bounds checks.
• Zero-length inputs no longer read element zero of an empty vector. fast_as_idate(), Inflate() and the internal Year()/YearMonthSplit() return empty results.
• fast_as_idate() no longer indexes before the start of its month table when the month component is zero or a month name is unrecognised (e.g. "01/00/2024", "01XYZ1948").
• Loops that call the R API (CHAR(), length()) are no longer run under OpenMP, as most of the R API is not thread-safe.
• Dates in December 2075, the last month of the supported range, are no longer reported as November: the binary search over the month table could not return its final element. format_idate() produced impossible strings such as "2075-11-61".

Bug fixes (dates)

• Years are now parsed from all four digits. "2999-01-01" was silently read as 1999; it is now rejected (or NaN when check = 0).
• A financial year now means the same thing to the input checker and to the converter. Previously, for fy_month >= 7, the two disagreed by one calendar year.
• An <fy> object now means the same thing as the equivalent "YYYY-YY" string. Previously it lost its class during conversion and was read as January of the ending year, whatever fy_month was.
• A quarter ("2015-Q1") is now the last month of that quarter everywhere. The checker used Mar/Jun/Sep/Dec and the converter Feb/May/Aug/Nov, which differ on a monthly index.
• The input checker now accepts exactly the strings the converter can read: a string of any other length (such as "2020"), or with arbitrary bytes in the separator positions, is rejected rather than passed and then silently converted to NaN.
• A fractional year (cpi_inflator(2015.9, 2016)) is an error rather than being truncated to 2015.

Bug fixes (fast_as_idate)

• format = "%d-%m-%Y" is now supported, as announced in 0.5.2. guess_format() could return it while the parser did not recognise it and read the date year-first instead.
• An unrecognised format is now an error rather than being read as "%Y-%m-%d".
• guess_format() now recognises day-first dates beginning 30 or 31, and reports the separator actually used.
• check now applies to every format, not only "%Y-%m-%d".

Bug fixes (custom series and forecasting)

• dr2index() no longer skips the first period after the end of the published series, and no longer adds a period beyond the requested date.
• .next_date() no longer produces a thirteenth month for a quarterly series anchored in October or November, and clamps the day of the month rather than producing an invalid date.
• r2index() no longer errors with "object 'next_date' not found".
• A trailing rate (cpi_original(2030, 0.1, 0.05)) is now dispatched correctly.
• Rates given as strings keep their sign and exponent: "-5%" was read as +5% and "1e-2" as 12. Rates are now validated against a strict grammar and a rate of -100% or less is rejected.
• The non-fable fallback extension built one more date than value for annual series (producing a recycling warning and a wrong final row), and used quarterly arithmetic for monthly series. Both are fixed.
• The fable path restores the data.table thread count with on.exit(), so a modelling error no longer leaves it changed.

Other changes

• All CPI, wage-price and labour-force series are now bundled with the package, so every inflator works immediately after installation without an internet connection. download_data() continues to store newer copies in the user data directory. A downloaded copy takes precedence only when its observation coverage is at least as recent as the bundled snapshot, so a cache left by an older package version cannot mask newer bundled data.
• index is now validated before use: it must have date and value columns, at least two rows, strictly increasing dates forming a regular annual, quarterly or monthly sequence, and finite non-zero values. The native code computes a row offset arithmetically, so an irregular or unsorted index would otherwise be read at the wrong position.
• A downloaded series is fully parsed and validated before it replaces the cached file, the replacement is a same-filesystem rename (with the previous version kept alongside as .bak), and the in-memory copy of a refreshed series is dropped so that download_data() takes effect within a session.
• GET_SERIES_FY() no longer adds an fy column by reference to the cached series.
• The test suite now runs the exported inflators on every platform, including machines without internet access, using the bundled ABS data.
• Removed a unit test that had been gated on a date now past, and so had stopped running.